civitai CLI

Browse and download Civitai models, images, and articles — and author, validate, and submit App Blocks. Two paths in one static binary: an anonymous read/download client for the public API, and the toolchain for shipping Apps.

⚠️ Apps is in a limited, invite-only beta (pre-GA). You can install this CLI, login, scaffold, validate, and run an app locally right now — but civitai app submit and dev:live require an invite: submission and dev:live are limited to invited beta testers while the feature is in a limited (pre-GA) beta, until Apps opens to the public.

Anyone can request an invite — open a request below and we'll review it:

The command-line interface for Civitai — a single static binary that does two things: it's a thin read/download client for Civitai's public API (browse and fetch models, images, and articles — no account needed to read), and it's the toolchain to author, validate, and ship Apps.

An App is a small, sandboxed web app that runs inside Civitai surfaces (it's served in an iframe; the platform owns the build and the runtime). The CLI replaces the error-prone "hand-format a ZIP" flow: it scaffolds a correct project, validates the manifest against the platform contract, and packages/submits it for review.

New here? The Build your first App guide is the full end-to-end walkthrough.

Install

Pick whichever fits — npm is the most convenient if you already have Node (App authors usually do); Homebrew is quickest on macOS/Linux; the prebuilt binary needs no toolchain; go install builds from source.

npm (Node)

A thin wrapper that downloads the matching prebuilt binary for your OS/arch on install and verifies its sha256 against the release checksums.txt:

npm install -g @civitai/cli
# or run it without installing:
npx @civitai/cli --help

Homebrew (macOS / Linux)

brew install civitai/tap/civitai

Nix flake

This repo is a Nix flake, so you can run or install civitai without a Go toolchain (works on x86_64/aarch64 Linux and macOS):

# Run without installing:
nix run github:civitai/cli -- models search "sdxl"

# Install into your Nix profile:
nix profile install github:civitai/cli

Pin it as an input in your own flake:

{
  inputs.civitai-cli.url = "github:civitai/cli";

  outputs = { self, nixpkgs, civitai-cli }: {
    # e.g. add to a devShell / home-manager / systemPackages:
    #   civitai-cli.packages.${system}.default
  };
}

Prebuilt binary

Download a prebuilt binary for your OS/arch from the GitHub Releases page (linux, macOS, windows × amd64/arm64), verify it against checksums.txt, then put it on your PATH:

tar xzf civitai_*_linux_amd64.tar.gz
sudo mv civitai /usr/local/bin/
civitai version

Go install (from source, Go 1.25+)

go install github.com/civitai/cli/cmd/civitai@latest
# installs the `civitai` binary into $(go env GOPATH)/bin

Quickstart: browse & download

Reads are anonymous — no login needed. Every command takes --json to emit the raw API response for scripting.

# Search models — filter by base model, type, and sort:
civitai models search --base-model Illustrious --type Checkpoint --sort "Most Downloaded"

# --base-model works on any type, including embeddings (TextualInversion):
civitai models search --type TextualInversion --base-model "SDXL 1.0"

# Inspect a specific model or a specific model version:
civitai models get 4384
civitai model-versions get 128713

# Download a version's file(s) — SHA256-verified, streamed atomically.
# `--layout` routes each file into the right app subfolder (also `a1111`);
# `--dry-run` prints the plan without transferring. Downloads require `civitai login`.
civitai download 128713 --layout comfyui --root ~/ComfyUI
civitai download 128713 --dry-run

# Find and read articles (guides) right in the terminal:
civitai articles search --query "comfyui workflow"
civitai articles get 32680 --content

See Browse the public API and Download model files below for the full command and flag reference (images, tags, creators, collections, pagination, folder routing, base-model compatibility checks, and more).

Quickstart: build an App Block

# 1. Authenticate once (browser device login; or `civitai login --token <t>`).
civitai login

# 2. Scaffold a ready-to-build App (batteries-included page-money default).
civitai app create my-app
cd my-app

# 3. Install deps and run it locally against the mock host (no real Buzz/compute).
#    `npm run dev` alone renders blank — the harness supplies the host.
npm install
npm run dev:harness

# 4. Edit your app; build it, then check the manifest before submitting.
#    (the `static` template has no build step — skip `npm run build`.)
npm run build
civitai app validate

# 5. Package + submit for review (uploads with your stored token by default).
civitai app submit

# 6. Check where your submission is in review / deploy.
civitai app status

Want to drive the real backend (real Buzz/compute) before submitting? Mint a dev token with civitai app dev-token and run npm run dev:live — see Local dev loop.

Submit ≠ live. civitai app submit enters your app into moderator review — it is not published immediately. The lifecycle is submit → review → approve → build + deploy → https://<blockId>.civit.ai/: that URL 404s until a moderator approves your submission and the platform builds + deploys it (a few minutes after approval). Until then, track status on /apps/my-submissions (a fresh submission sits at pending). See Submit & auth for the full flow. (And note Apps is in an invite-only beta — see the warning above.)

Enable shell completion (optional):

source <(civitai completion bash)   # bash; see `civitai completion --help` for zsh/fish/powershell

SDK packages

This CLI scaffolds, validates, and submits — but the code your app actually imports lives in two published npm packages (the page-money template wires them for you; static and page-vite are deliberately dependency-free):

Package	What it is
@civitai/blocks-react	The React hooks + iframe transport app authors call — useBlockContext, useBuzzWorkflow, useBlockResize, the /ui component pack, and the /testing dev hosts. Start here for the hook reference.
@civitai/app-sdk	The framework-agnostic contract under the hooks — manifest types, scope strings, the postMessage protocol, and the defineBlock validator (@civitai/app-sdk/blocks).

# Already installed by the scaffold; this is the explicit install line:
pnpm add @civitai/blocks-react @civitai/app-sdk react

The full hook-by-hook reference (with snippets) lives in each package's npm README. For the end-to-end walkthrough, see Build your first App.

Command reference

Command	What it does
civitai login [--token [<t>]] [--no-browser]	Browser OAuth device login by default (stores auto-refreshing tokens); --token <t> stores a personal API key instead. --token with no value prints where to create a personal key (civitai.com/user/account) and how to re-run — handy when you know you want a personal key but haven't minted one yet. Config at ~/.config/civitai/config.yaml, 0600. Also reads CIVITAI_TOKEN.
civitai whoami [--scopes] [--json]	Verify the stored token; print the authenticated user and a Capabilities section — credential type (OAuth login vs personal API key), Read Buzz balance, and Spend Buzz — decoded from the token's scope, so a money-path dead end (OAuth login can't spend) is visible before dev:live. --scopes also lists every granted scope; --json emits the user + credentialType/canReadBalance/canSpend/scopes (scriptable).
civitai buzz [--json]	Show your spendable Buzz balance (blue / green / yellow, plus a total). Needs a full-scope personal API key to read; an OAuth login token can't, and gets a clear "switch to a personal key" message. --json emits {blue,green,yellow,total} (scriptable — handy for before/after diffing a dev:live spend).
civitai app create [name] [dir] [--template static|page-vite|page-money] [--dir <path>] [--name <display>]	The friendly happy path. Scaffold a ready-to-build App, defaulting to the batteries-included page-money SDK template (default dir ./<slug>).
civitai app init [name] [dir] [...]	Same scaffolder as create with a no-build static default (back-compat alias).
civitai app dev-token <slug> [--env]	Mint a short-lived (~4h) dev block token for npm run dev:live — calls the invite-gated mint route with your stored credential, reading scopes from your local block.manifest.json (so it works on an unsubmitted slug). Prints the token (--env prints VITE_LIVE_BLOCK_TOKEN=<token>, paste-ready); warns at mint time if the token is read-only (can't spend). See Local dev loop.
civitai app dev-tunnel [blockId] [--port] [--tunnel-endpoint] [--idle-timeout]	(Pre-GA / dark) Preview your local dev server inside the real Civitai host at civitai.com/apps/dev/<blockId> — a prod-fidelity inner-dev-loop. Mints an ephemeral in-memory ssh keypair, opens a reverse tunnel from your dev port (start npm run dev:tunnel first) to the Civitai tunnel endpoint, prints the URL to open, and tears everything down on Ctrl-C or an idle timeout. Before minting it also pre-flights whether the host can actually embed your dev server — the host iframes it sandboxed (opaque null origin), so a dev server missing Access-Control-Allow-Origin: *, missing the .civit.ai entry in allowedHosts, or sending a framing header that excludes civitai.com loads as a blank iframe with no error anywhere. Those are printed as warnings (never fatal) just above the URL, with the vite.config.ts fix. Apps scaffolded by civitai app init --template page-money already satisfy all of it. Gated behind an Apps-author invite and a kill-switch flag that is off today, and the tunnel endpoint is not exposed yet — so it reports "not available" until it ships.
civitai app validate [dir] [--strict] [--json]	Best-effort local pre-check of block.manifest.json; emits non-fatal warnings (--strict fails on them). --json emits the structured result (ok, plus errors/warnings each with field/message) for scriptable parsing — still exits non-zero on failure. See Validate fidelity.
civitai app submit [dir] [--package-only] [--out f.zip] [--skip-validate]	Validate + package the source tree + upload it with your stored token (or, with no token, write the bundle + print next steps).
civitai app status [blockId] [--id <pubreq>] [--json]	Check the review/deploy status of your own submissions. No arg lists them all; a blockId (app slug) or --id shows one in detail (rejection reason if rejected, live URL once deployed). See Submission status.
civitai app metrics <slug> [--from <d>] [--to <d>] [--json]	Owner-only analytics for one of your Apps — installs, runs + Buzz spent, Buzz purchased, and API engagement. Always prints the window the server served (it defaults to 30 days and clamps to 366), so a zero is never ambiguous. Needs a personal API key (an OAuth login is refused). See App metrics.
civitai app withdraw [pubreq-id] [--id <pubreq>]	Withdraw your own pending submission (the pubreq_… id from civitai app status). Frees the slug so a fresh civitai app submit can replace it. Idempotent; only a pending request can be withdrawn. See Submission status.
civitai generate "<prompt>" [--negative-prompt <p>] [--quantity <n>] [--aspect-ratio <r>] [--checkpoint <version-id>] [--lora <version-id>[:strength]] [--image <path-or-url>] [--ecosystem <key>] [--input <file>] [--print-input] [--dry-run] [--json] [--max-cost <buzz>] [--yes] [--no-wait] [--timeout <dur>] [--out-dir <dir>] [--no-download] [--force] [--external-id <key>]	Generate images from a text prompt — this SPENDS REAL BUZZ. Prices the job with the server's estimator, shows the cost + your balance, asks before spending, submits, then waits and downloads the results into --out-dir as <workflow-id>-<n>.<ext>. --no-wait prints the workflow id and exits; --timeout bounds the wait (never the job and never the charge); --no-download waits but prints URLs instead of writing files. --dry-run estimates and exits without submitting (--dry-run --json emits the raw estimate). --print-input prints the assembled graph and exits without reaching any money seam (no submit, no estimate, no balance read) — note that with --image it still uploads local files first, because the printed graph has to reference real blob URLs for --input to be able to submit it; uploading spends nothing; --input <file> (or - for stdin) sends a raw graph as-is — txt2img only, and mutually exclusive with the content flags. --image <path-or-url> (repeatable) attaches a reference image for image-to-image — a local png/jpeg is uploaded, an https URL is passed through — and requires --ecosystem, because without one the server ignores the images, generates from the prompt alone and charges anyway. Needs a personal API key with the AI Services scopes; an OAuth login is refused. --max-cost is an estimate check, not a spending cap. See Generate.
civitai workflows list [--limit <n>] [--cursor <c>] [--tag <t>] [--json]	List the generation workflows you have submitted, newest first — status, when, cost, and deliverable/total outputs. Cursor-paged: the next cursor is printed on stdout when more results exist. Reading spends nothing. See Generate.
civitai workflows get <workflow-id> [--json]	Look up one generation workflow — status, steps and outputs. This is how you re-attach after --no-wait, a --timeout expiry or a Ctrl-C. Outputs that are blocked, unavailable or hidden are listed with the reason rather than omitted. Output URLs are presigned and expire; re-run for fresh links. Reading spends nothing. See Generate.
civitai workflows cancel <workflow-id> [--yes] [--json]	Stop a running generation. 🔴 This does not refund anything — a mid-run cancel bills the accrued cost, non-refundably. Cancel because you no longer want the output, never to save money. Asks for confirmation (default no); --yes skips the prompt and a non-TTY without it refuses. See Generate.
civitai version	Print version / commit / build date.
civitai completion [shell]	Generate a shell-completion script.

Run civitai help, civitai app --help, or civitai <command> --help for the full details and examples.

Templates

• static — a no-build page app (index.html + a tiny app.js, block.manifest.json with page:{}, no build step).
• page-vite — a Vite + React page app with config-as-code build fields (buildCommand: "npm run build" + outputDir: "dist").
• page-money — a Vite + React + TypeScript full-page (W10) money-path app wired to the published App SDK (@civitai/blocks-react + @civitai/app-sdk): prompt → estimate → lazy consent → submit → poll → real Buzz spend, via useBuzzWorkflow / useRequestConsent / useBlockResize (never raw postMessage). Ships a dev:harness mock host, .env.* allowed parent-origin config, and a unit-test stub. Run npm run dev:harness (plain npm run dev renders blank without a host).

The host handshake (BLOCK_READY)

Every template declares a page surface, and the host will not reveal a page app until the app posts BLOCK_READY — that handler is the only transition into the host's ready state. An app that never sends it is replaced by a visible failure card once the host's bounded retries run out, even though the app itself renders perfectly. Nothing you can run locally reproduces that.

page-money gets the handshake for free: @civitai/blocks-react's iframe transport acks internally, which is why the SDK templates never touch raw postMessage. The two SDK-free templates (static, page-vite) therefore ship a small vendored emitter, civitai-host.js, loaded from the entry point. Leave it in place.

⚠️ If you adopt @civitai/blocks-react, delete civitai-host.js in the same change. This is the one situation where removing it is correct, and running both is worse than running neither: whichever handshake answers the host's first BLOCK_INIT cancels the host's retry loop and its readiness timeout. If the vendored emitter wins that race, the SDK transport can be left never having seen an init — its waitForInit rejects after 10s and the host sits "ready", showing an app that never started, with no retry and no error card.

Two rules it encodes, which apply to every message you add afterwards:

• The envelope is { type, payload }. The host dispatches event.data.payload to its subscribers, so fields put at the top level ({ type: 'X', height: 0 }) arrive as payload: undefined.
• Answer, don't announce. The ack goes out in response to the host's BLOCK_INIT, addressed at the origin that init arrived from rather than broadcast to '*'. It is also why nothing is posted when you preview locally: there is no host to send BLOCK_INIT, so the emitter stays silent by design.

🔒 The emitter checks the sender window, not the sender's identity. It answers window.parent — whoever framed you — which is sound for this one message because the ack carries no data. It is not sufficient for anything you add next. The moment you handle an inbound message carrying a token, a viewer, storage or a result, check event.origin against an allowlist of origins you trust, or any page that frames your app can feed it whatever it likes. The emitter deliberately does not vendor that allowlist — the real list (production, preview subdomains, dev tunnels) is platform state that moves without notice, and @civitai/blocks-react already maintains it from VITE_BLOCK_ALLOWED_PARENT_ORIGINS. Adopt the SDK before you handle data.

RESIZE_IFRAME is not part of a page app's protocol: the host renders a page block full-viewport, so it does not size to content and ignores the message. (useBlockResize is surface-agnostic and page-money still calls it — on a page surface it is simply a no-op, which is why the SDK templates can share component code across surfaces.) The pre-#206 templates demoed a raw postMessage of RESIZE_IFRAME, so a project scaffolded before that fix still carries dead code you can delete. (This CLI's own CI fails if a shipped template ever reintroduces it; there is no author-facing command that scans your project for it — civitai app validate checks the manifest and the handshake, not this.)

Local dev loop (harness: mock vs live)

A scaffolded App is a sandboxed iframe, and locally there is no host to send BLOCK_INIT — so npm run dev shows you your own UI and nothing of the protocol. The page-money template ships a dev harness (the SDK's @civitai/blocks-react/testing hosts) to close that gap, with two modes:

Command	Mode	What it does
npm run dev:harness	mock (default)	Mounts the SDK mock host — synthetic replies, no real Buzz, no compute, no network. Safe to spam; drive money/error/insufficient-Buzz UX via on-screen scenarios or ? URL params. Start here.
npm run dev:live	live	Mounts the SDK live host (createLiveHost) — forwards the App protocol to the real Civitai backend with a pasted dev token (Bearer). Spends REAL Buzz / real compute.

⚠️ dev:live works on a pending (un-approved) app. The dev-token mint (POST /api/v1/blocks/dev-token) accepts a pending slug — right after a successful civitai app submit (status pending) it returns 200 with appId: pending-pubreq_… and dev:live mounts the live host against the pending app. For real generation you must mint with a full-scope personal API key; an OAuth (civitai login) token mints read-only (user:read:self) and cannot spend. Use civitai buzz / civitai whoami to confirm your credential can spend before a live run.

Live mode needs a short-lived dev block token. Mint it with civitai app dev-token (the CLI handles the invite-gated POST /api/v1/blocks/dev-token call with your stored credential — no hand-rolled curl) and paste it into .env.development.local as VITE_LIVE_BLOCK_TOKEN=:

# From your scaffolded project dir (reads scopes from block.manifest.json):
civitai app dev-token my-block --env >> .env.development.local
npm run dev:live

.env.development* is never committed (submit excludes it) and the token is short-lived (~4h) — re-run dev-token when it expires. Mint with a full-scope personal API key for real generation; an OAuth login mints a read-only token (the command warns you at mint time). With no token, dev:live fails safe (renders a notice, never spends). Live v1 covers the money path (estimate/submit/poll/cancel); pickers, checkpoint-set, App-Storage KV, and in-band Buzz purchase are mock-only.

Under the hood (the scaffold wires this — you don't configure it): dev:live routes the live host's backend calls through the vite dev proxy (server.proxy['/api']), not straight to civitai.com: createLiveHost fetches /api/... SAME-ORIGIN against the dev server (localhost:5186), and vite proxies that server-side to civitai with the Origin header rewritten to an allowlisted host. This is load-bearing — a direct cross-origin fetch from localhost is both blocked by CORS preflight and rejected by civitai's tRPC origin gate. The same-origin proxy + Origin rewrite fixes both. VITE_LIVE_HOST_ORIGIN overrides the proxy target (default https://civitai.com).

Which credential can spend? Only a full-scope personal API key can spend Buzz — whether that is a real dev:live generation in your app or a civitai generate run from the terminal. The default OAuth login can't do either:

Credential	Can spend Buzz? (dev:live, civitai generate)	How to get it
Personal API key (full scope)	✅ Yes — estimate → submit → generation → real Buzz	create it in the web UI at civitai.com/user/account, then civitai login --token <key> (a personal key carries AI Services)
civitai login (OAuth, default)	❌ No — viewer + catalog + app storage only	the civitai-cli client has no AI Services scope, so the server strips the spend scope — fine for read/identity dev:live, not for generation

This is the single most common blocker for civitai generate: an OAuth login looks perfectly valid, and the refusal is a scope problem, not a login problem — re-running civitai login will not fix it. civitai whoami shows the capability as Spend Buzz (AI Services).

You can't mint a personal key over OAuth or the CLI (apiKey.add returns 403 without a full-scope session) — create it in the web UI. The dev token always grants user:read:self, so your viewer resolves on either path. For the scope mechanics behind this, see Submit & auth.

Env vars (VITE_BLOCK_ALLOWED_PARENT_ORIGINS, VITE_HARNESS_MODE, VITE_LIVE_BLOCK_TOKEN, …) and the scenario knobs are documented in depth in the scaffolded project's own README.md and .env.example — see internal/scaffold/templates/page-money/README.md.tmpl.

Examples

Two real example manifests live under examples/ (copied from the civitai-block-* dogfood apps). Read them for manifest shape — between them they cover the required fields, $schema wiring, the page/iframe blocks, and scope declarations with justifications:

• examples/buzz-generator.block.manifest.json
• examples/notepad.block.manifest.json

The values are those apps' own choices, not recommendations. In particular don't copy buzz-generator's page.buzzBudgetPerGen — it is a safety ceiling against a malicious or compromised app, not an estimate of one run, so size your own from the field's description in the canonical schema (notepad doesn't take the budgeted scope, so it has no budget at all).

Both validate clean (examples_test.go asserts this so the claim stays true) — schema conformance only, which says nothing about whether a value is well-sized.

Browse the public API

Beyond authoring Apps, the CLI is a thin client for Civitai's public read REST API (GET /api/v1/**). These subcommands work anonymously — no login needed, because the data is public — but when you're logged in your stored token is sent automatically (pass --anon to force a no-auth request). Every command also takes --json to print the raw API JSON response for scripting.

Command	What it does	Notable flags
civitai models search	Search models (GET /api/v1/models)	--query, --tag, --username, --type, --base-model (repeatable), --sort, --period, --nsfw; paging --limit (≤100), --page, --cursor
civitai models get <id>	Get one model by id	--json, --anon
civitai model-versions get <id>	Get a model version by id (alias mv)	--json, --anon
civitai model-versions by-hash <hash>	Look up a model version by file hash (AutoV2, SHA256, …)	--json, --anon
civitai download <version-id>	Download a model version's file(s)	--model, --file, --all, --out, --out-dir, --layout, --root, --for-base, --no-verify, --force, --anon
civitai images search	Search images (GET /api/v1/images)	--model-id, --model-version-id, --post-id, --username, --base-model (repeatable), --type (image/video/audio), --sort, --period, --nsfw, --meta (include generation metadata); paging --limit (≤200), --page, --cursor
civitai tags search	Search model tags	--query; paging --limit (≤200), --page
civitai creators search	Search creators	--query; paging --limit (≤200), --page
civitai users get <username-or-id>	Look up a user via public search (a number = exact id; a name = exact-username match, else it lists close matches)	--json, --anon
civitai articles search	Search articles (GET /api/v1/articles)	--query, --tags, --username, --sort, --nsfw; paging --limit (≤100), --cursor
civitai articles get <id>	Get one article by id (--content renders the article body as readable text/markdown)	--content, --json, --anon
civitai collections search	Search public collections (GET /api/v1/collections)	--query, --sort, --nsfw; paging --limit (≤100), --cursor
civitai collections get <id>	Get one collection by id	--json, --anon

Pagination. List commands print a compact footer with the next-page hint. models/images support both shallow --page and deep --cursor paging (the API caps page*limit at 1000 and 429s beyond it — prefer --cursor for deep paging); articles/collections are cursor-only (keyset feed — no --page); tags/creators are --page-only. Each endpoint caps --limit (models/articles/collections 100; images/tags/creators 200).

civitai models search --query "pony" --limit 5
civitai models get 4384
civitai model-versions by-hash 5D8D26E2A6
civitai articles get 32680
civitai articles get 32680 --content   # render the article body (the guide) as readable text/markdown
civitai images search --model-id 4384 --sort "Most Reactions" --json   # raw JSON for scripting

Filtering by base model. --base-model is repeatable and maps to the REST baseModels filter (an OR across the values). It's the key discovery filter for things --type can't separate — e.g. video checkpoints all share --type Checkpoint and are distinguished only by base model. It works on both models search and images search:

civitai models search --type Checkpoint --base-model "Wan Video 2.2 T2V-A14B"
civitai models search --base-model Pony --base-model Illustrious --limit 20
# images too — find recent-popular images generated with a given base model:
civitai images search --base-model "Krea 2" --sort "Most Reactions" --period Week
civitai images search --type video --sort "Most Reactions"   # videos only

Generation metadata (--meta). By default the image list is a compact table without generation data (matching the API, which omits meta unless asked). Add --meta to include each image's prompt, sampler, cfg, steps, seed, and model — rendered as an indented detail block per image (the table can't hold a prompt). Images whose uploader chose to hide their generation data show meta: (hidden by uploader). With --json, --meta adds the raw meta object to each item.

civitai images search --nsfw --sort "Most Reactions" --period Month --meta
civitai images search --model-version-id 128713 --meta --json | jq '.items[].meta'

The human table includes a BASE MODEL column (the base model each image was generated with, when the API reports one; - when it doesn't), so you can see the ecosystem at a glance without dropping to --json.

--sort is ignored with --model-id. The REST API returns images for a given modelId in its own default order regardless of sort, so images search --model-id <id> --sort … prints a one-line note on stderr and the results are NOT re-sorted. (--model-version-id is unaffected — it honours --sort.)

Non-weights file marker. In the human (non---json) output of models get and model-versions get, a version whose primary file is not model weights (type != "Model") is tagged with its actual file type — e.g. [Archive] (a "Workflows" model's downloadable deliverable), [Training Data], or [Other] — so you can see at a glance that the version's file isn't weights. It's purely informational: any file type still downloads. --json output is an unchanged raw passthrough.

Download model files

civitai download fetches the file(s) of a model version. Identify the version deterministically by its numeric version id, or resolve a model's default (first) published version with --model:

civitai download 128713                       # the version's primary file → ./<server-name>
civitai download --model 4384                 # resolve model 4384's default version, then download its primary file
civitai download --model 4384 --dry-run       # print the plan (files, sizes, hashes, targets) — download nothing
civitai download 128713 --out ./dreamshaper.safetensors
civitai download 128713 --file vae --out-dir ./models   # pick a file by name; write into a dir
civitai download 691639 --file 1234567                  # pick one of two same-named files by its file id
civitai download 128713 --all --out-dir ./models        # every file in the version
civitai download 128713 --all --layout comfyui --root ~/ComfyUI   # route each file to its type folder
civitai download 128713 --layout a1111 --for-base "SDXL 1.0"      # A1111 layout + base-model compat warning

Downloads require authentication. Every model-file download needs a token — even a small public embedding 401s anonymously. Run civitai login first. The read/search commands work anonymously; downloads do not. --anon is meaningful for the read commands, not for download.

Behavior:

• Identifier — exactly one of the positional <version-id> or --model <model-id> is required (no numeric-ambiguity guessing).

• --model resolves the default version — the model's default (first published) version; its primary file is downloaded regardless of file type. Any model type works, including a type: Workflows model whose deliverable is a downloadable Archive.

• --dry-run — resolve the version + selected file(s) and print the plan (each file's name, size, SHA256, resolved target path, and whether authentication will be required) then exit 0, transferring nothing and creating no file (not even a .part). Works with --file, --all, --model, --out, --out-dir, and --layout/--root (the plan shows the routed target paths).

• File selection — defaults to the version's primary file. --file selects one file by numeric file id (the version's files[].id) or by name (exact, else a unique case-insensitive substring; ambiguous/none errors and lists the candidate files with their ids). --all downloads every file.

• Same-named files (no silent overwrite) — a version can ship two files that share a name (e.g. Flux Dev's fp16 and fp8, both flux_dev.safetensors). Selecting that shared name with --file is ambiguous and errors, listing both files with their ids — pass the numeric id to pick exactly one (--file 1234567; the id is shown by --dry-run and in the error). --all refuses to run when two selected files would resolve to the same on-disk path (which would silently clobber one) — it fails before transferring anything, lists the colliding files with their ids/sizes, and tells you to pick one with --file <id> (or write them to separate paths). No download ever silently overwrites another.

• Output — --out <path> sets an exact target path (single file only). --out-dir <dir> writes server-named files into a directory (works with --all). Parent directories are created as needed. Default is the server-provided filename in the current directory.

• Type-aware folder routing (--layout) — --layout <a1111|comfyui> writes each file into the correct subfolder for that app, keyed by the file/model type, under --root <dir> (default .). This fixes the footgun where --all --out-dir X dumps a bundled VAE into the checkpoint folder and pollutes the model dropdown: with --layout, the checkpoint lands in the checkpoints folder and the VAE in the VAE folder. --layout is mutually exclusive with --out/--out-dir; --root only applies with --layout. An unmapped type (Poses, Wildcards, Archive, …) is written to --root with a stderr note rather than silently misplaced. The routed folder maps:

  Civitai type	A1111 / Forge	ComfyUI
  Checkpoint	models/Stable-diffusion	models/checkpoints
  VAE (standalone or bundled)	models/VAE	models/vae
  LORA / LoCon / DoRA	models/Lora	models/loras
  TextualInversion (embedding)	embeddings	models/embeddings
  Hypernetwork	models/hypernetworks	models/hypernetworks
  Controlnet	models/ControlNet	models/controlnet
  Upscaler	models/ESRGAN	models/upscale_models

  (Sources: the AUTOMATIC1111 wiki + the sd-webui-controlnet models/ControlNet default; the ComfyUI models docs.)

• Mis-file warning (without --layout) — when --all would place files of differing types into one directory (the mis-file footgun), the CLI prints a one-line stderr warning naming the off-type file(s) and suggesting --layout. It's a warning, not an error; a single-type download stays quiet.

• Base model + compatibility (--for-base) — the version's base model is always shown in the plan/output. --for-base "<baseModel>" warns on stderr when the version's base model is in a confidently different family than your target (e.g. an SD 1.5 embedding like EasyNegative downloaded for an SDXL 1.0 model → won't work; the wrong VAE → black images). The check is conservative — it groups the common bases into architecture families (SD1.x, SD2.x, the SDXL family [SDXL/Pony/Illustrious/NoobAI, treated loosely], SD3, Flux, video, …) and only warns on an architecture-level mismatch, never on near-neighbours (Pony vs Illustrious) or unclassifiable bases.

• Streaming + atomicity — the body streams to <target>.part and is renamed into place only on success, so an interrupted run never leaves a truncated final file. Large files (10+ GB) are never buffered in memory. TTY-aware progress is printed to stderr. The Civitai download URL 302-redirects to signed storage; the CLI follows it.

• Auth — your stored login token (civitai login) or CIVITAI_TOKEN is used automatically; Civitai requires a token to download any model file, even public ones, so an anonymous download gets an actionable 401 (401 → run civitai login; 403 → the file is gated for your account). --anon forces no token.

• Transient-failure retry (reads) — the read endpoints (search / model / version / images / tags / creators / users / articles / collections) retry a transient 502/503/504 or network error a few times with exponential backoff (with jitter), noting each retry on stderr. A 429 is retried only when it carries a Retry-After header (a genuine throttle, honored up to a cap); a 429 without Retry-After is Civitai's deterministic deep-paging limit and is surfaced immediately with the hint to use --cursor instead of --page. The download stream is not retried mid-transfer.

• Integrity (default on) — the streamed bytes are verified against the file's SHA256; a mismatch deletes the .part and fails. --no-verify skips it; a file with no published SHA256 downloads with a warning (not a hard failure). Note that SHA256 verifies integrity (the bytes match what the API advertised), not authenticity — it proves the download wasn't corrupted or truncated in transit, but a compromised source that advertises a matching hash for malicious bytes cannot be detected by the hash alone. Only download models from creators you trust.

• Pickle/archive safety note — when a downloaded file has a pickle/executable extension (.ckpt, .pt, .pth, .bin, .pickle, .pkl) or an archive extension (.zip, .tar, .tar.gz, .tgz, .rar, .7z), the CLI prints a one-line stderr note: these formats can execute arbitrary code when loaded by ComfyUI/A1111/torch.load, and they land in folders those apps auto-scan. safetensors and image files are inert and get no note. The note is informational — it never blocks the download.

• ControlNet preprocessor note — when the parent model is a ControlNet, the CLI prints a one-line stderr note: a ControlNet model needs a matching preprocessor/annotator (e.g. the ComfyUI comfyui_controlnet_aux custom node — OpenPose/Canny/Depth) to derive the control image from your input, and that preprocessor is a separate install, not hosted on Civitai. The note is informational — it never blocks the download.

• Idempotency — an already-present target (that verifies, or with --no-verify) is skipped with a note; --force re-downloads.

• Any file type downloads — the selected/primary file is downloaded whatever its type (Model weights, a type: Workflows model's Archive, training data, or other artifacts). The human models get / model-versions get output tags a non-weights primary file with its type (e.g. [Archive]) purely for information; it never blocks a download.

Scripting with --json

Every read subcommand takes --json, which prints the raw /api/v1/... REST response — a stable passthrough, not a CLI-invented shape. So the field schema is exactly the public Site API's; keep the REST field reference open (e.g. models, model-versions) rather than reverse-engineering fields with jq keys.

Two properties make the output safe to pipe:

• --json stdout is pure JSON — nothing else is written to stdout, so ... --json | jq -e . always parses.
• Errors go to stderr with a non-zero exit — a failed call writes the error to stderr, exits non-zero, and prints nothing to stdout, so jq never sees error prose. For example civitai model-versions get 999999999 --json exits 4 with Error: not found (404): Model not found on stderr and an empty stdout.

Both properties hold for civitai generate and civitai workflows … too, but their payloads are not Site API REST shapes — generation has no REST route, so those commands pass through the raw orchestrator reply. Read Generation --json before scripting against them.

Cursor pagination loop

For deep paging use --cursor (not --page — the API caps page*limit at 1000 and 429s beyond it). Read .metadata.nextCursor from each response and feed it back via --cursor; stop when it's absent/null:

export CIVITAI_NO_UPDATE_CHECK=1
cursor=""
while :; do
  page=$(civitai models search --type LORA --base-model Illustrious \
           --sort "Most Downloaded" --limit 5 ${cursor:+--cursor "$cursor"} --json) || break
  echo "$page" | jq -r '.items[].id'                 # do your work here
  cursor=$(echo "$page" | jq -r '.metadata.nextCursor // empty')
  [ -z "$cursor" ] && break                          # no more pages
done

Clean output for pipelines

The CLI runs a background check for a newer release and prints a nag to stderr. In scripts, silence it with CIVITAI_NO_UPDATE_CHECK=1 (env) or --no-update-check (flag). Either way stdout stays pure JSON — the nag never touches stdout — but suppressing it keeps stderr clean for logs.

Generation --json

civitai generate --dry-run --json, civitai workflows list --json and civitai workflows get <id> --json emit the raw orchestrator payload. Two caveats have bitten people, and neither shows up as an error:

• Output URLs are presigned and EXPIRE. The links in a workflow payload are short-lived signatures, not durable addresses. A pipeline that stores them and fetches later gets a 401/403 from the storage host that no credential can fix — re-run civitai workflows get <id> for fresh links instead of caching the old ones. (Fetch them with no Authorization header; they are already authorized, and the CLI deliberately attaches nothing to them.)

• --json still exits 0 when the job is not generatable. --dry-run --json prints the estimate and exits 0 even when the payload says "ready": false, which means the server has already decided it cannot serve this job (an unavailable or unsupported resource). A human --dry-run prints a warning and a real submit refuses outright, but a script reading only the exit code sees success. Branch on the field, exactly as app metrics requires branching on notOwned:

  q=$(civitai generate "a cat" --checkpoint 128713 --dry-run --json) || exit $?
  [ "$(echo "$q" | jq -r .ready)" = "true" ] || { echo "not generatable" >&2; exit 1; }
  echo "$q" | jq -r .cost.total

  Cost keys (cost.factors, cost.fixed) are server-owned and passed through verbatim, so treat them as an open map rather than a fixed set.

Gotchas

• SHA256 is UPPER-case in the API/--json (e.g. 42BA94DF20CC0F4E6DF46E3C294587A2F8CF133BF0134185884EE1C9C5E108C4), while sha256sum emits lowercase. Case-fold before comparing if you roll your own verify (civitai download's built-in check is already case-insensitive): [ "$(echo "$api_sha" | tr A-Z a-z)" = "$(sha256sum file | cut -d' ' -f1)" ].
• models search already embeds .modelVersions[] — each item carries its full versions, including files[].hashes.SHA256 and trainedWords. If you're iterating search results you usually don't need a follow-up model-versions get per version.
• Creator + model-level download counts live only in the search response. model-versions get <id> returns a version, whose .model is just {name, type, nsfw, poi} — no creator, no model stats.downloadCount. If you started from a version and need those, fetch them from models search / models get and join on the model id (.modelId on the version).

Worked example — top LoRAs for a base model, then plan a download

Search → pick versions with jq → hand each version id to download with app folder routing. --dry-run prints the plan (files, sizes, hashes, target paths) without transferring, so this snippet is safe to copy-paste:

export CIVITAI_NO_UPDATE_CHECK=1
civitai models search --type LORA --base-model Illustrious \
    --sort "Most Downloaded" --limit 3 --json |
  jq -r '.items[].modelVersions[0].id' |
  while read -r vid; do
    civitai download "$vid" --layout comfyui --root ~/ComfyUI --dry-run
  done

Drop --dry-run (and civitai login first) to actually fetch the files — --layout comfyui routes each into its ComfyUI type folder.

Validate fidelity

civitai app validate is a best-effort LOCAL mirror of the platform's approve-time validator (BlockManifestValidator). The server is the source of truth at review time — passing validate locally is a strong pre-check, not a guarantee of approval.

It checks block.manifest.json against a vendored JSON Schema (schema/app-block.manifest.schema.json, syntactic shape) plus the ported semantic rules the server runs (sandbox trust-tier allowlist, page ⇒ iframe, required iframe sub-fields, the renderMode tier gate, targets[].slotId registry membership) and structural project checks. A few checks are necessarily approximate locally (the slot registry is vendored; per-app origin-binding/scope checks the CLI can't see are not reproduced).

It also mirrors one build-time rule, because the failure it prevents is otherwise an opaque server-side "build failed": your committed lockfile must match the package manager the platform derives from buildCommand. The platform build installs strictly from the lockfile — no registry re-resolve fallback — so "buildCommand": "pnpm run build" needs pnpm-lock.yaml, "yarn run build" needs yarn.lock, and npm run … / vite build / npx vite build / an omitted buildCommand all need package-lock.json. A mismatch or a missing lockfile is a hard validate error; an extra unused lockfile is a warning. Apps with no package.json are static — the platform never installs for them and they are never flagged.

Finally it emits one advisory about the host handshake: if your manifest declares a page surface and nothing in your source posts BLOCK_READY, validate says so. That is the shape of an app scaffolded before the templates were fixed (#206) — it renders perfectly everywhere you can look locally and is replaced by a failure card in the real host. It is a warning, never an error: unlike the lockfile rule (where the platform build provably dies), this one infers runtime behaviour from static text and can be wrong, so it must not fail a correct project. Three things follow:

• A dependency that acks ends the check — today that is @civitai/blocks-react, and nothing else. Its iframe transport acks internally and the literal never appears in your src/, so a page-money app is never flagged. This is an exact list, not the @civitai/ scope: @civitai/app-sdk is the server-side SDK and no runtime code in it posts BLOCK_READY, and @civitai/theme / @civitai/components are CSS. Depending on those does not give you the handshake, so it does not silence the check either.
• It reads source only — never node_modules, never the conventional build directories (dist, build, out, …), and never a .md file: a README describing the handshake is not an implementation of it. Comments are stripped too, so a comment naming BLOCK_READY does not satisfy it. A src that is a symlink into a shared package is followed.
• It stays quiet when it cannot see the whole project. An unreadable file, a file over 2 MiB, a very large tree, or a directory holding only a manifest all mean "we could not look" — reported as nothing, never as a finding.

If it fires on a project you know is correct — your ack arrives from a bundled dependency, or from a file type this scan doesn't open — it is a false alarm, and it never blocks (exit 0) unless you pass --strict. What it proves is narrow: that the message is mentioned in code. It cannot prove the ack ever fires; only the real host can.

civitai app submit prints the same warnings before it uploads, and likewise does not block on them.

⚠️ If you already run civitai app validate --strict in CI, this advisory is new and can turn a previously-green project red — which is what --strict asks for. If it is a false alarm for your project, drop --strict or add the ack, and please open an issue: a warning at a correct project is a bug in the check, not something you should have to work around.

The durable fix is a server-side civitai app validate endpoint that calls the real BlockManifestValidator (the faithful contract), with this schema published as the syntactic half. See AGENTS.md for the full caveat and how the vendored schema + Go checks are kept in sync.

Submit & auth

civitai login (no flags) runs the OAuth device-authorization grant: it prints a URL + a short code, you approve in your browser, and the CLI stores a short-lived access token (1h) plus a refresh token (30d) that it rotates automatically before requests and once on a 401. It requests the scopes UserRead | AppBlocksSubmit (== 33554433, exactly the civitai-cli OAuth client's allowedScopes) — identity plus Apps submit, which gates both app submit and the dev-token mint. It deliberately does not request AIServicesWrite: the server's device-flow scope check is all-or-nothing, so asking for a scope the client doesn't allow would reject the whole login. A login token therefore drives the read/identity dev:live paths (viewer, catalog, app storage) but — for a generation app whose only ai:write:budgeted scope is stripped — cannot estimate, submit, or spend real Buzz. For real generation use a full-scope personal API key (see the credential table under Local dev loop above), which carries AI Services. civitai login --token <key> stores a personal API key instead (no refresh). CIVITAI_TOKEN overrides the stored credential (treated as a personal key).

civitai app submit:

• always validates + packages the canonical source ZIP, then
• uploads it with your stored token to the token-authenticated route POST /api/v1/blocks/submit-version (Authorization: Bearer). OAuth tokens refresh transparently. Set CIVITAI_SUBMIT_PATH to override the route.
• With no token configured (and not --package-only), it instead writes the .zip and prints the next steps (civitai login, or web upload at /apps/submit).

--package-only always just writes the .zip and stops.

After you submit: review → approve → deploy

A successful submit does not publish your app — it queues it for moderator review. The lifecycle is:

1. submit → your submission lands at /apps/my-submissions with status pending.
2. review → a moderator reviews the manifest + files. They either approve or reject (with a reason you can read inline, then fix and resubmit).
3. deploy → on approval, the platform builds and deploys your app (injects its build recipe → builds the image → deploys → programs the <blockId>.civit.ai DNS record). A few minutes after approval it serves live at https://<blockId>.civit.ai/.

Before approval, https://<blockId>.civit.ai/ 404s — submitting does not make the subdomain serve (but dev:live works against a pending app — see Local dev loop). For the full end-to-end walkthrough (build → submit → review → deploy), see the Build your first App guide.

Need to change the bundle while a request is still pending? Withdraw it first to free the slug, then resubmit:

$ civitai app status                          # find the pubreq_ id
$ civitai app withdraw pubreq_01HZX           # frees the slug
$ civitai app submit                          # resubmit the new bundle

civitai app withdraw <pubreq-id> (or --id <pubreq>) withdraws your own pending publish request. It is idempotent (an already-withdrawn request still returns success) and only a pending request can be withdrawn — an already approved/rejected one cannot.

Submission status

civitai app status checks where your own submissions are in that lifecycle without leaving the terminal. It calls the token-authenticated, self-scoped route GET /api/v1/blocks/submissions with your stored credential — you only ever see your own submissions (the same token that submitted can read its status; OAuth tokens need the Apps submit scope).

With no argument it lists every submission, newest first:

$ civitai app status
BLOCK_ID    VERSION  STATUS    DEPLOY    SUBMITTED   URL
gen-matrix  0.6.0    approved  live      2026-06-22  https://gen-matrix.civit.ai/
my-block    0.2.0    pending   -         2026-06-21  -
old-app     0.1.0    approved  building  2026-06-19  -

Pass a blockId (app slug) or --id <pubreq_id> to see one in detail — including the rejection reason if it was rejected (so you can fix + resubmit) and the live URL once it is approved and deployed:

$ civitai app status gen-matrix
Block ID:         gen-matrix
Version:          0.6.0
Publish request:  pubreq_01HZX
Status:           rejected
Deploy state:     -
Submitted:        2026-06-22 09:05 CDT
Reviewed:         2026-06-22 11:40 CDT

Rejection reason:
  the budgeted scope needs the per-app Sybil cap signed off first

Not live yet — gen-matrix.civit.ai only serves after the app is approved and deployed (deployState 'live').

The unfiltered listing is capped server-side at 100 rows, and the API returns no cursor and no total — so there is no way to page and no way to know how many were dropped. When a full-length page comes back the CLI says so on stderr rather than presenting it as your complete history:

note: showing the newest 100 submissions — the API caps this listing and offers no way to page, so older submissions may exist but are not listed. Look up a specific app with `civitai app status <blockId>`.

That is an inference (a page that is exactly full is indistinguishable from one that was cut off), so it says may. A per-app lookup — civitai app status <blockId> — is not affected: the server narrows to the slug before applying the cap.

--json emits the raw response for scripting. An empty list prints a friendly "run civitai app submit" hint; with no token it points you at civitai login. Notes like the cap caveat go to stderr, so --json stdout stays pure and the exit code stays 0.

App metrics

civitai app metrics <slug> shows the owner-only analytics for one of your App Blocks. The slug is resolved to its appBlockId through your own submissions, so analytics exist only once a version has been approved — an app still in review reports that instead of an empty dashboard.

$ civitai app metrics gen-matrix --from 2026-05-01 --to 2026-08-03
App:          gen-matrix
Window:       2026-05-01 00:00 UTC → 2026-08-03 00:00 UTC
Granularity:  week

Installs
  Total   12
  Active  9

Runs
  Count       20
  Buzz spent  65

Buzz purchased
  Purchases  3
  Buzz       15000
  Gross      $14.97

App loads
  Impressions       124
  Unique viewers    12
  Signed-out loads  40

Engagement
  API calls     26
  Active users  2
  Error rate    3.8%

  Top scopes:
    ai:write:budgeted  20

  Top endpoints:
    /api/v1/blocks/me  4

Three things are worth knowing, because each one otherwise produces a believable-but-wrong reading:

• The window is always printed, and it comes from the server. The API defaults to the last 30 days and clamps any request to 366 days, so a real app with 20 runs in mid-June reads 0 under the default window. The window shown is the one the server actually served — if it clamped your --from, the printed range says so. Widen it with --from / --to, which accept a plain YYYY-MM-DD (midnight UTC) or a full RFC3339 timestamp; a malformed value or an inverted window is a usage error (exit 2) caught before any request.
• "Not entitled" is not "zero". When the caller doesn't own the app (or lacks Apps-author access) the API answers HTTP 200 with every counter zeroed, flagged only by a notOwned field. The CLI refuses to render a dashboard in that case and tells you to check civitai whoami / civitai app status <slug> instead — a silently-empty dashboard that looks like real data is the failure mode this command is built to avoid.
• It needs a personal API key. The query is full-scope, so an OAuth civitai login token gets a 403; the error names the fix (civitai login --token <key>).

Two data caveats.

Engagement counts only authenticated, scope-gated API calls. An app that ships no scoped API surface shows real installs and revenue with a flat engagement section — that is expected, not a bug. Installs is a different case again: it shows n/a for an app that cannot be installed at all (a page app has no install slot, so an install record cannot exist), which is deliberately distinct from a real 0 on an installable app nobody has installed yet. App loads is the exception: it is measured on every load, so it counts signed-out visitors and static blocks that engagement structurally cannot see. Unique viewers counts signed-in people once each and approximates signed-out ones by network address, so read it as reach rather than an identity count, and Signed-out loads is a count of LOADS (one anonymous visitor reloading ten times is 10 there and 1 unique viewer), so it can legitimately exceed Unique viewers. Note also that these are mount ATTEMPTS: a load that FAILED still counts, because a failed mount's only beacon is the same event and it carries no status. Error rate is the share of those calls that failed, and the human view renders it as a percentage (the server sends it as a 0–1 ratio, which --json passes through unchanged). Only a genuine zero prints 0.0%: a real but tiny rate — a high-traffic app with a handful of failures — reads <0.1% rather than rounding away to look error-free.

--json emits the raw analytics payload (the server's own object, including notOwned and the per-bucket series arrays the human view omits) for scripting. Note that --json does not refuse a not-entitled read the way the human view does: a notOwned: true payload is passed through with every counter zeroed and the command still exits 0, so civitai app metrics <slug> --json | jq .runs.count returns 0 for an app you can't see. A script must branch on the notOwned field rather than trusting the counts.

Installs carries installs.notApplicable for the case above. It is NOT an outage flag — it means the question does not apply to this app type, so a script should render it as "not applicable" rather than retrying or warning about infrastructure. --json passes it through and still exits 0, so branch on it rather than trusting the counts.

App loads has a SECOND, section-local unavailability flag — views.unavailable, independent of notOwned. It is the one section the server reads from a different store, which can be unreadable — or merely too slow, the read is time-bounded server-side — while every other counter in the same response is genuinely measured. When that happens the human view prints unavailable and says so explicitly rather than printing a 0 you would read as "nobody opened my app". --json passes the flag through and still exits 0, so a script must branch on views.unavailable too — jq .views.count alone cannot tell an outage from a real zero. A server old enough to predate this section omits the views key entirely; the human view reports that as unavailable as well (naming the different cause), and a script should treat a missing .views the same way.

Generate

civitai generate "<prompt>" runs a text-to-image generation on Civitai's generator.

🔴 This spends real Buzz and cannot be undone. A submitted generation is charged. There is no cancel-for-refund. Price it with --dry-run first — that calls the cost estimator and spends nothing.

# Price it. Spends nothing.
civitai generate "a cat wearing sunglasses" --dry-run

# The same estimate as raw JSON, for scripts
civitai generate "a cat wearing sunglasses" --dry-run --json

# Generate, refusing if the estimate exceeds 50 Buzz
civitai generate "a cat wearing sunglasses" --quantity 4 --max-cost 50

# A specific checkpoint plus a LoRA at 0.8 strength
civitai generate "a cat" --checkpoint 128713 --lora 250712:0.8

# Wait for the result and write the images into ./out
civitai generate "a cat" --yes --out-dir ./out

# Fire and forget; collect the results later
civitai generate "a cat" --yes --no-wait
civitai workflows list
civitai workflows get <workflow-id>

# Non-interactive (CI) — --yes is required, or the run is refused
civitai generate "a cat" --yes --max-cost 20

# Image-to-image from a local file — --ecosystem is REQUIRED with --image
civitai generate "make it winter" --ecosystem Flux1Kontext --image ./cat.png --dry-run

# …or from a public https URL, with two reference images
civitai generate "combine these" --ecosystem Seedream \
  --image https://example.com/a.jpg --image ./b.png --yes

# Graduate from flags to a raw graph: print, edit, send back
civitai generate "a cat" --quantity 2 --print-input > graph.json
civitai generate --input graph.json --dry-run

Credential. Generation needs a full-scope personal API key carrying the AI Services scopes (create one, then civitai login --token <key>). An OAuth browser login (civitai login) does not carry them and is refused. civitai whoami shows the capability as Spend Buzz (AI Services).

🔴 --max-cost is an estimate check, not a spending cap

The cost this command shows is an estimate, not a quote: the server's estimator returns no quote id, no signed price and no expiry, so there is nothing to hand back at submit time — and no server-side spending ceiling is reachable from an API key at all. The realized charge can exceed the estimate and is not refunded.

--max-cost compares that estimate against your number and refuses locally before submitting. It catches a --quantity typo. That is all it can do. Do not run an unattended loop believing it caps spend. (The per-API-key buzzLimit on your account does not bind this path either — the generator meters a separate server-minted subject, not your key.)

Confirmation

An interactive run prints the estimate, your balance and the resolved model names, then asks. A non-interactive shell (pipe/CI) without --yes is refused rather than charged silently. Everything the confirmation prints goes to stderr, so --json keeps stdout machine-clean.

Image-to-image: --image and --ecosystem

--image <path-or-url> (repeatable) attaches a reference image and turns the job into an edit. A local .png/.jpg is uploaded to Civitai first and the stored blob is referenced; an https URL is passed through as-is, but must be publicly reachable — the generator downloads it server-side too, and an unfetchable URL is a 400 after you have already been priced. Either way the CLI reads the image's width and height from its header only (never decoding the pixels) and sends them, because the server requires both and rejects an entry without them. http://, file:// and data: are refused, local files are capped at 64 MiB (checked by stat, before a byte is read), and only png and jpeg are supported — webp would need a new third-party decoder dependency.

🔴 --image requires --ecosystem, and the reason is money. The server promotes a text-to-image job to image-to-image only when the request names an ecosystem. Without one it ignores the images, generates from the prompt alone, and charges you the full amount — HTTP 200, no error, no warning. Measured: the same graph with and without images[] priced byte-identically.

Two more things the CLI genuinely cannot check for you, so it says them instead of pretending:

• Only some ecosystems accept reference images at all. Qwen, Flux1Kontext, NanoBanana, Seedream, OpenAI, Grok, Reve, MAI, Boogu and a few more do; the Stable Diffusion family and the default ecosystem do not — and for those the images are dropped silently and billed. The cost estimate cannot tell you which case you are in: several edit-capable ecosystems price identically with and without images (measured on Flux1Kontext, NanoBanana and Seedream), so a price comparison is not a detector. Name an ecosystem you know supports editing.
• Too many reference images are silently truncated. Per-ecosystem limits run from 1 to 7, live only inside the server's per-engine graphs, and the extras are dropped before any limit check can fire — so the server never reports it and the truncated job is billed. Measured on Qwen (limit 3): 4, 5, 6 and 12 images all priced identically to 3. The CLI refuses more than 7 (no ecosystem accepts more, so that refusal can never block a valid request) and warns for anything above 1. It deliberately does not vendor the per-ecosystem table — see AGENTS.md items 13 and 18.

--ecosystem is sent to the server verbatim and is not checked locally; an unknown value comes back as the server's own unknown ecosystem error.

--dry-run does upload local --image files, because an estimate built on a graph with no images[] prices a plain text-to-image job. Uploading spends no Buzz, and --dry-run still never submits.

The content flags, and why there aren't twelve

--negative-prompt, --quantity, --aspect-ratio, --checkpoint <version-id>, --lora <version-id>[:strength] (repeatable), plus --image / --ecosystem above.

The generator is permissive, not a validator — it returns HTTP 200 for things it silently changes:

• An out-of-range --quantity is clamped with no error (asking for 40 charges you for the server's limit). The CLI warns when you cross it.
• --steps 0 / --cfg-scale 0 are accepted and price a degenerate, cheaper, wrong job — which is exactly why those flags are not exposed yet.
• A checkpoint id that does not exist is accepted, the ecosystem default is silently substituted, and you are billed for it.

So --checkpoint and every --lora is resolved against the public model-version API before anything is submitted: a bad id becomes a hard local not found (exit 4) instead of a wrong charge, and the confirmation echoes the resolved model name so you approve a name rather than an integer. --model is deliberately absent: civitai download --model takes a model id, while this takes a version id.

Raw graphs: --print-input and --input

The five flags cover the common job. Everything else the generator understands lives in the generation graph — the JSON document the flags assemble. You can write that document yourself:

# 1. Assemble it from flags, print it, and exit. No submit, no cost estimate,
#    no balance read — with no --checkpoint/--lora, no request at all.
civitai generate "a cat" --quantity 2 --aspect-ratio 1:1 --print-input > graph.json

# 2. Edit graph.json however you like.

# 3. Send it as-is. Price it first; --dry-run still spends nothing.
civitai generate --input graph.json --dry-run
civitai generate --input graph.json --yes

# …or pipe it, with `-`
jq '.prompt = "a dog"' graph.json | civitai generate --input - --dry-run

--print-input reaches no money seam: not the submit, not the cost estimator, not the balance read. With --checkpoint/--lora it does still make the public model-version read those flags always make — that lookup supplies model.type, which graph resources[] require, so skipping it would print a document --input could not submit.

--print-input's output is a valid --input document by construction — that round-trip is the point of the pair, and it is what replaces a --set some.path=value expression language the CLI deliberately does not have (a wrong type in such an expression is accepted by the server silently, and billed; an edited file is inspectable before it is sent).

Four things to know, all of them consequences of it being a passthrough:

• txt2img only. A graph declaring any other workflow is refused. The server's content audit reads the top-level prompt node, and it rebuilds what it inspects from declared graph nodes — so a graph carrying its prompt somewhere else (a comfy node, a nested step input) is exactly the shape that could reach the generator unaudited. That question is open upstream, and this CLI will not be the path that answers it the wrong way.
• Envelope keys are refused, not ignored. civitaiTip, creatorTip, buzzType, tags, externalId, sourceMetadata, sourceMetadataMap, remixOfId and a top-level input belong to the request envelope around the graph, not to the graph. A file setting civitaiTip would charge a tip that --dry-run structurally cannot show you — the estimator prices a strictly smaller request and is never sent tips at all — so the file is rejected with an error rather than quietly cleaned up.
• Keys the CLI does not model are passed through, with a warning. The warning says the CLI cannot verify the key; it is not a claim that the key is invalid, because the CLI does not carry a copy of the server's node registry. It matters because the server's failure mode for a key it does not declare is to drop it silently at HTTP 200 — a typo costs Buzz and produces a job that ran without your parameter, with no error anywhere.
• No model-id safety net. --checkpoint / --lora are resolved against the public API before submitting; a raw graph is not interpreted, so a nonexistent id in it is accepted, the ecosystem default is substituted, and you are billed.

--input cannot be combined with a prompt argument or with --negative-prompt / --quantity / --aspect-ratio / --checkpoint / --lora — there is no predictable answer to "does --lora append to or replace the file's resources?", so the combination is a usage error. Every execution flag (--dry-run, --yes, --max-cost, --json, --no-wait, --timeout, --out-dir, --no-download, --force, --external-id) still applies.

Waiting, downloading, and re-attaching

By default generate waits for the job to finish and writes every deliverable output into --out-dir (default .) as <workflow-id>-<n>.<ext>. --force overwrites existing files; without it a collision is refused before any bytes move.

• --no-wait submits, prints the workflow id and exits 0.
• --no-download waits and prints the output URLs instead of writing files.
• civitai workflows get <workflow-id> shows a workflow at any time. It is the re-attach path for every case where the CLI stopped early, and it spends nothing.

🔴 --timeout stops waiting. It does not stop paying. When the deadline passes (or you press Ctrl-C) the generation keeps running server-side and the charge stands — there is no cancel-for-refund, and a mid-run cancel bills the accrued cost anyway. Both cases exit non-zero, print the workflow id, the idempotency key and the exact civitai workflows get … command, and never report success.

Output URLs are presigned and expire. Download promptly; re-read the workflow for fresh links. The blob fetch deliberately carries no credential — the URL is already authorized, and attaching your full-scope API key to it would hand 25 unrelated permissions to a request that needs none.

A finished workflow can contain fewer usable results than you paid for. An output can be blocked by moderation, never land, or be one you hid on the website. Those are filtered out of the download — and reported, with the reason, plus an explicit note when the count differs from --quantity. Silently writing three files for a four-image job is the failure this exists to prevent. If every output is filtered out the command exits non-zero.

Crash safety. The orchestrator's idempotency key is written to ~/.config/civitai/pending/<key>.json before the request is sent, because the money moves server-side even if the process dies mid-POST. If a submit's reply never arrives, re-run with --external-id <key>: the orchestrator dedupes on it and returns the pre-existing workflow instead of charging again (it answers a duplicate with HTTP 200, not a 409, so re-attachment is inferred locally).

Polling cadence. The status poll starts at 5s, backs off exponentially to a cap, and backs off harder on a 429. That floor is not tunable downward: the workflow read proxies straight through to the orchestrator with no cache and no server-side rate limit, so the CLI's own restraint is the only thing between it and a 429 storm.

Listing and cancelling workflows

civitai workflows list                      # newest first
civitai workflows list --limit 5
civitai workflows list --limit 50 --cursor <next-cursor>
civitai workflows list --json               # raw server payload, incl. nextCursor
civitai workflows cancel <workflow-id>      # asks for confirmation
civitai workflows cancel <workflow-id> -y   # skip the prompt (scripts/CI)

list is cursor-paged, not page-numbered: when more results exist it prints Next cursor: <c> on stdout, which you pass back as --cursor. --tag filters on orchestrator workflow tags (repeatable).

The OUTPUTS column reads deliverable/total. The two differ when an output was blocked by moderation, never landed, or you hid it on the website — so 0/4 means four images were produced and paid for and none of them are usable, which is a very different fact from 0/0. civitai workflows get <id> shows the per-output reason.

🔴 cancel does not refund anything. A mid-run cancel bills the accrued cost, orchestrator-side and non-refundably. There is no cancel-for-refund on this platform: by the time a workflow is running the money has moved. Cancel a job because you no longer want its output — never as a way to save Buzz, and never to undo a submit. (This is also why --timeout and Ctrl-C deliberately do not cancel: stopping the wait costs nothing, while stopping the job would cost the same as letting it finish and throw the result away.)

cancel asks for confirmation, matching civitai generate and civitai app submit. It is the one irreversible action here, and it destroys a job you have already paid for, so it is gated the same way every other destructive path in this feature is:

• --yes / -y proceeds without prompting;
• an interactive terminal prints what is lost and prompts — the default is no, so a bare Enter aborts;
• a non-interactive shell without --yes refuses rather than cancelling silently. Scripts must pass --yes explicitly.

Nothing is cancelled when the confirmation is refused — the gate runs before the request goes out.

Exit codes specific to generate

generate follows the global exit-code table, with one deliberate refinement. The API answers several very different failures with the same HTTP status, and the generic mapping would send a script down the wrong path — in particular a caller who is out of Buzz, muted, or hitting a server-side outage must never be told to re-run civitai login. Those cases therefore exit 1 (generic), not 3 (auth) or 2 (usage):

🔴 An exit code does not tell you whether you were charged. Every failure above the divider happens before anything is submitted, so nothing was spent. Every failure below it happens after the submit, and the Buzz is gone — including a --timeout, a Ctrl-C, and a workflow that ends failed. Do not write a retry loop that branches on the exit code alone; re-attach with civitai workflows get <workflow-id> instead of re-submitting.

Failure	Exit
— nothing submitted, nothing spent —
Missing AI Services scope / no token / not authenticated	3
Not enough Buzz (caught locally against your balance, or reported by the server)	1
Account muted, or onboarding incomplete	1
Generation disabled server-side	1
Prompt refused by content moderation — 🔴 never retry, repeated blocked prompts get the account muted	1
The server priced the job but reports ready: false (a selected resource is not currently generatable)	2
Estimate above --max-cost, an unknown ecosystem, or a resource that resolved fine but is "not enabled for generation" (the ids exist; the combination is not runnable — distinct from exit 4, which means "no such id")	2
--input that is malformed, declares a non-txt2img workflow, carries an envelope key (civitaiTip, …), or is combined with a content flag	2
No such --checkpoint / --lora version id	4
civitai workflows get / workflows cancel on an unknown workflow id (a read; spends nothing)	4
— 🔴 submitted: the Buzz is already spent —
--timeout expired, or Ctrl-C while waiting — the job keeps running server-side and was not cancelled	1
The workflow finished failed / expired / canceled	1
The workflow succeeded but every output was filtered out (blocked / unavailable / hidden)	1

Configuration

Setting	Config key	Env var	Default
Personal API key	token	CIVITAI_TOKEN	—
OAuth tokens (device login)	auth_kind, access_token, refresh_token, token_expiry, scope	—	—
API base URL	base_url	CIVITAI_BASE_URL	https://civitai.com
Submit endpoint	—	CIVITAI_SUBMIT_PATH	/api/v1/blocks/submit-version

Config lives at ~/.config/civitai/config.yaml (honours XDG_CONFIG_HOME), written owner-readable only.

Exit codes

civitai returns a differentiated exit code so scripts can branch on the kind of failure without parsing stderr. The human-readable error message is unchanged by this — only echo $? differs.

Code	Meaning
0	Success.
1	Generic / unclassified error.
2	Usage error — a bad flag, a bad flag value (e.g. --limit out of range, a non-integer id), or a request the API rejected as malformed (HTTP 400, e.g. a bad --period/--sort enum).
3	Authentication/authorization — login required, token invalid/expired, or the credential lacks the needed scope (HTTP 401/403, or no token configured). civitai generate refines this: several of its failures are not credential problems but would otherwise land here or on 2, so they exit 1 instead and a script never loops on civitai login. A muted account or incomplete onboarding arrives as a bare 403 that is byte-identical to a missing scope; out of Buzz and generation disabled arrive as 400 (the upstream 403 is re-thrown server-side as a tRPC BAD_REQUEST), which would otherwise read as "bad flags". See Generate.
4	Not found — the requested resource does not exist (HTTP 404).
5	Network/transport failure or service unavailable — dial/timeout, or HTTP 502/503/504 after retries.
6	Rate limited — throttled by the API (HTTP 429).

# Branch on failure kind
if ! civitai models get "$id" >/dev/null 2>&1; then
  case $? in
    3) echo "log in first: civitai login" ;;
    4) echo "no such model: $id" ;;
    5|6) echo "transient — retry later" ;;
    *) echo "failed" ;;
  esac
fi

Troubleshooting

• no token configured — run civitai login (or set CIVITAI_TOKEN).
• unauthorized (401) — your token is invalid/expired. OAuth tokens refresh automatically; if the refresh token has also expired, run civitai login again. For a personal key, create a new one at https://civitai.com/user/account and civitai login --token <key>.
• forbidden (403) / service unavailable (503) — your account may lack Apps access while the feature is in its invite-only beta (see the warning at the top of this README). Submission is limited to invited beta testers until Apps reaches general availability.
• validation failed — read each - ... line; fix the manifest, or pass --skip-validate to package anyway (the server will still re-validate).
• <dir> is not empty — refusing to overwrite — app init won't clobber an existing directory; pick a new name or remove the directory.

Development

make ci      # go mod tidy + vet + test + build (mirrors CI)
make test
make build   # -> bin/civitai
make fmt
go test ./... -cover

• Language: Go 1.25, Cobra (commands) + Viper (config).
• Layout / conventions / how to add a command / release process: see AGENTS.md.
• Contributing: see CONTRIBUTING.md.

CI (.github/workflows/ci.yml) runs go vet, gofmt -s -l ., go test ./..., and go build ./... on every push/PR.

Releasing

Releases are built by goreleaser from a GitHub Actions workflow on a v* tag push:

git tag v0.1.0
git push origin v0.1.0

This cross-compiles for linux/darwin/windows × amd64/arm64, stamps version/commit/date, and publishes a GitHub Release with archives + checksums.txt plus a Homebrew tap bump. See AGENTS.md for the full process and the secrets it needs (HOMEBREW_TAP_GITHUB_TOKEN).

License

Apache License 2.0.