Spindle

MCP server for multi-harness AI agent delegation. Spawn background agents (Claude Code, Codex, Gemini, Kimi) that run asynchronously, with optional git worktree isolation for safe parallel work.

Features

• Async agent spawning - Fire-and-forget pattern with spool IDs
• Optional blocking with gather/yield - Wait for all results at once, or stream them as agents complete. Alternatively, agent can continue other work, spins are nonblocking by default
• Permission profiles - Control what tools child agents can use (readonly, careful, full)
• Shard isolation - Run agents in sandboxed git worktrees to prevent conflicts
• Model selection - Route tasks to different models per-agent
• Session continuity - Resume conversations while the upstream provider session remains available; saved transcripts remain available for manual reconstruction when it does not
• Rich querying - Search, filter, peek at running output, export results
• spindle doctor - One command to check the install: versions, paths, spool store, detected harnesses, and an optional live read-only smoke

Requirements

• Python 3.11+
• At least one harness CLI, installed and already authenticated. Spindle shells out to the CLI you already use and inherits its login — it never asks for an API key of its own:
  • Claude CLI (claude) — the default harness
  • Codex CLI (codex)
  • Gemini CLI (gemini)
  • Kimi CLI (kimi-cli)
• Git (for shard/worktree functionality)
• bwrap (bubblewrap), on Linux, for filesystem-contained shards and every non-full Kimi launch

Install

pip install spindle-mcp

Check what the install can actually see and do:

spindle doctor

Doctor reports this CLI's version and path, whether the service on the port is this same install, whether the spool store is writable, and which harness CLIs it found. Add --smoke to have it spawn one real read-only headless agent per harness and verify the answer comes back:

spindle doctor --smoke

Add to Claude Code's MCP config (~/.claude.json):

{
  "mcpServers": {
    "spindle": {
      "command": "spindle"
    }
  }
}

That runs spindle over stdio, which needs no background service. Run a service only if you want the HTTP transport or a shared long-lived instance — see Background service.

Command line

The core spawn-and-collect loop is available as subcommands, so you can drive spindle from a shell or a script without an MCP client. (The querying tools — search, grep, stats, export, and the shard commands — are MCP-only for now.) Commands print JSON by default; --human prints text.

# Spawn an agent (returns a spool id immediately)
spindle spin "Summarize this module" --harness codex --permission readonly -d /path/to/project

# Any built-in harness or lodged profile name works
spindle spin "Quick pass" --harness claude-code --model haiku
spindle spin "Alt endpoint" --harness my-profile

# Collect
spindle unspool <spool_id>
spindle spools --human
spindle peek <spool_id> -n 100
spindle wait <id1>,<id2> --mode yield
spindle drop <spool_id>

# Settle custody only when doctor reports an abandoned owner episode
spindle repair <spool_id> --attest-dead

# This install
spindle --version
spindle doctor [--smoke] [--json] [--port N]
spindle status [--port N]

spindle spin takes the same harness names as the spin tool: the four built-ins plus any lodged profile. An unknown name is an error, not a silent fallback to Claude Code.

Usage

Basic: Spawn and collect

# Spawn an agent
spool_id = spin("Research the Python GIL")

# Do other work...

# Check result
result = unspool(spool_id)

Permission profiles

Control what tools the spawned agent can use:

# Read-only / manual: the one tight, no-exec tier (allowlist-enforced)
spin("Analyze the codebase", permission="readonly")   # "manual" is an alias

# Careful (default): classifier-vetted auto — CC vets each tool call server-side
spin("Fix this bug", permission="careful")

# Full access: for initial setup, dependency installs, environment provisioning
spin("Set up a new Python project with dependencies", permission="full")

# Shard: Full access + auto-isolated worktree (common for risky work)
spin("Refactor the auth system", permission="shard")

# Careful + shard: classifier-vetted, isolated in a bwrap-contained worktree
spin("Update configs", permission="careful+shard")

# Research: web/file research routed to a SKEIN site, a single file, or a directory
spin("Research deepseek vs kimi", permission="research", research_target="site:spindle-development")

Profiles (claude-code harness):

• readonly (alias manual): Read, Grep, Glob, safe bash (ls, cat, git status/log/diff). The only tier still governed by an allowlist — no python, no find, no write. This is the tight, inspectable, manual option.
• careful (default): now an alias of auto. No allowlist; runs under --permission-mode auto, where Claude Code vets each tool call server-side on intent. Use it for most code work including reviews/fells. (It used to be a Bash allowlist that gated capability on command phrasing, not security — auto removes that gate.)
• full: No restrictions
• shard: Full access + auto-creates isolated worktree (bypass inside the bwrap-contained shard)
• careful+shard: auto semantics + auto-creates isolated worktree (bypass inside the bwrap-contained shard)
• research: Read, Grep, Glob, WebFetch, WebSearch, curl, jq, safe bash; no python/find; requires research_target (Write/Edit added when target is file: or dir:)
• research+shard: research tools + auto-creates isolated worktree

Web-egress work (WebFetch, WebSearch, curl) belongs in research — the other code tiers intentionally have no web access so they're safe for code review and code-modifying work.

You can also pass explicit allowed_tools to override the profile.

Isolated workspaces with shards

Run agents in isolated git worktrees to prevent conflicts:

# Agent works in its own worktree
spool_id = spin("Refactor auth module", shard=True)

# Check shard status
shard_status(spool_id)

# Merge changes back when done
shard_merge(spool_id)

# Or discard if not needed
shard_abandon(spool_id)

Shards create a git worktree + branch. If SKEIN is available, uses skein shard spawn for richer tracking. Falls back to plain git worktree otherwise.

Wait for completion

# Spawn multiple agents
id1 = spin("Find all TODO comments")
id2 = spin("List unused imports")
id3 = spin("Check for type errors")

# Gather: block until all complete, get all results
results = spin_wait("id1,id2,id3", mode="gather")

# Yield: return as each completes
# Great when results are independent - process each as it lands
result = spin_wait("id1,id2,id3", mode="yield")  # Returns first to finish

# With timeout
results = spin_wait("id1,id2", mode="gather", timeout=300)

Yield mode keeps you responsive instead of blocking on the slowest agent.

Time-based waiting

Simple timed waiting with spin_sleep:

spin_sleep("90m")       # Sleep for 90 minutes
spin_sleep("2h")        # Sleep for 2 hours
spin_sleep("30s")       # Sleep for 30 seconds
spin_sleep("06:00")     # Wait until 6 AM

Or use spin_wait with the time parameter:

spin_wait(time="90m")
spin_wait(time="06:00")  # Handles next-day wraparound

Useful for periodic check-in loops (e.g., QM/dancing partner patterns).

Model selection and timeouts

# Route quick tasks to haiku (fast, cheap)
spin("Summarize this file", model="haiku")

# Complex work to opus
spin("Design the new architecture", model="opus")

# Frontier reasoning: "fable" is the CLI's rolling alias (Fable 5.1 as of 2026-09)
spin("Untangle the scheduler race", model="fable")

# Pin a generation instead of riding the rolling alias
spin("Implement the new architecture", model="opus-5.5")   # claude-opus-5-5
spin("Reproduce last month's review", model="fable-5")      # legacy claude-fable-5

# Auto-kill if it takes too long
spin("Should be quick", timeout=60)

Continue a session

# Get session ID from completed spool
result = unspool(spool_id)  # includes session_id

# Continue that conversation
new_id = respin(session_id, "Follow up question")

If Claude reports that the provider session is unavailable or expired, the respin terminates with that provider error. The saved transcript remains available for manual reconstruction in a new conversation.

Cancel running work

spin_drop(spool_id)

List all spools

spools()

Large results

Most results are small and return whole. Very long results (over ~50K chars) are truncated by unspool() to their head and tail, with a breadcrumb showing how to retrieve the rest. The full text always stays in the spool; truncation only shapes the default read.

# Default read - budgeted (head + tail if the result is huge)
unspool(spool_id)

# Get the entire result, no truncation
unspool(spool_id, full=True)

# Page through a slice
unspool(spool_id, offset=12000, limit=20000)

# Write the full result to a file (agent-driven, not automatic)
spool_export(spool_id, format="md", output_path="/tmp/result.md")

Tune the thresholds with SPINDLE_UNSPOOL_MAX_CHARS (default 50000), SPINDLE_UNSPOOL_HEAD_CHARS (12000), and SPINDLE_UNSPOOL_TAIL_CHARS (12000).

Search and filter

# Search prompts and results
spool_search("authentication")

# Filter by status and time
spool_results(status="error", since="1h")

# Regex search across all spool results
spool_grep("error|failed|exception")

# Dig into one huge result - matching lines with context, no full pull
spool_grep("error|failed", spool_id="abc123", context=3)

# Get statistics
spool_stats()

# Export to file
spool_export("all", format="md")

Multi-Harness Support

Spindle supports multiple AI agent harnesses, allowing you to choose the best tool for each task.

Available Harnesses

Claude Code (default) - Anthropic's Claude models via claude CLI

• Superior code understanding and reasoning
• Best for complex refactoring, architecture decisions
• Slower startup (~3-4 minutes to first response)
• Use harness="claude-code" or omit harness parameter

Codex CLI - OpenAI models via codex CLI

• Extremely fast startup (~10 seconds to first response)
• Good for quick edits, simple tasks, prototyping
• Requires ChatGPT Plus/Pro/Enterprise
• Models include "astra" (gpt-6-astra), "sol" (default), "terra", "luna", "reserve" (gpt-reserve, fast/affordable coding tier), or any full model name
• Use harness="codex"

Gemini CLI - Google's Gemini models via gemini CLI

• Fast startup (~5-10 seconds to first response)
• Full agent with tool use, file access, multi-step reasoning
• Generous free tier (1000 req/day with Google account)
• Models: "flash"/"video" (3.8 Flash), "pro" (3.1 Pro Preview, default), "flash-lite" (3.5 Flash-Lite), or any full model name
• Native video: put @clip.mp4 in the prompt and set working_dir to its directory. Verified with Gemini CLI 0.39.1; local files are limited to 20 MiB. See video setup and limits.
• Use harness="gemini"

Kimi CLI - Moonshot AI's Kimi models via kimi-cli

• Fast startup (~5-10 seconds to first response)
• Thinking mode for complex reasoning
• Models: "k3"/"latest"/"thinking" (K3, always thinking; default), "k2.7-code", "k2.6", or any full model name
• Kimi headless mode auto-approves its tool calls; it has no classifier-vetted or genuinely "careful" mode. Spindle therefore wraps every Kimi process except full without shard intent in its own bwrap filesystem boundary. The default makes only the requested working directory and Kimi's state directory writable, with private /tmp and filesystem-path runtime sockets hidden behind a private /run; host processes and IPC are in separate namespaces. The network namespace remains shared, so localhost services and abstract Unix sockets remain reachable. Shards substitute their worktree, and research adds its explicit output target. External Git metadata stays read-only, so Kimi leaves shard changes uncommitted. The caller inspects and commits them in the worktree, then calls shard_merge. Kimi refuses to start with a read-only work directory, so shared names such as readonly and careful do not narrow its powers inside that box. Shard intent always stays contained; full without a shard explicitly runs uncontained. Shard launches require an absolute KIMI_SHARE_DIR when that override is used.
• Use harness="kimi"

Basic Usage

# Claude Code (default) - best for complex work
spool_id = spin("Refactor the auth module to use dependency injection")

# Codex CLI - fast for simple tasks
spool_id = spin(prompt="Add error handling to this function", harness="codex", working_dir="/path/to/project")

# Gemini CLI - fast with free tier
spool_id = spin(prompt="Summarize this codebase", harness="gemini", working_dir="/path/to/project")

# Kimi CLI - fast reasoning with thinking mode
spool_id = spin(prompt="Analyze this bug", harness="kimi", working_dir="/path/to/project")

# All harnesses use the same API
result = unspool(spool_id)  # Auto-detects harness

Choosing a Harness

Use Claude Code when:

• Task requires deep reasoning or architecture decisions
• Working on complex refactoring across multiple files
• Need thorough code review or analysis

Use Codex when:

• Need quick edits or simple implementations
• Prototyping or exploring ideas rapidly

Use Gemini when:

• Want fast results without API key management (Google account login)
• Running many parallel tasks on a budget (free tier)
• Need a quick general-purpose agent

Use Kimi when:

• Need thinking mode for complex reasoning at speed
• Want fast startup with strong reasoning capabilities

Requirements

Claude Code:

• Claude CLI installed and authenticated

Codex CLI:

• Codex CLI installed (npm i -g @openai/codex)
• ChatGPT Plus/Pro/Enterprise subscription

Gemini CLI:

• Gemini CLI installed (npm i -g @google/gemini-cli)
• Google account login (gemini → "Login with Google") or GEMINI_API_KEY env var

Kimi CLI:

• Kimi CLI installed (pip install kimi-cli)
• Auth via kimi-cli login or API key in ~/.kimi/config.toml
• Bubblewrap (bwrap) installed for normal use; only permission="full" opts out of Spindle's external filesystem containment

See docs/MULTI_HARNESS_GUIDE.md and docs/CODEX_SETUP.md for detailed documentation.

Profiles

A profile is a named, lodged configuration: a base harness plus a set of overrides (model, alt-endpoint env, extra CLI flags). The motivating use is running any Anthropic-compatible model through the existing Claude Code harness by injecting ANTHROPIC_BASE_URL / ANTHROPIC_API_KEY / CLAUDE_CONFIG_DIR into the spawned child — spindle's output parsing, unspool, and respin all work unchanged because the child is still plain Claude Code.

A profile is a folder (the folder name is the profile name) containing a single profile.json. Profiles are discovered from two locations, later overriding earlier:

1. Canonical: ~/.spindle/profiles/<name>/profile.json — where real, private profiles live, outside any repo.
2. Dev convenience: ./profiles/<name>/profile.json relative to the current working directory (gitignored).

Use a profile by passing its name as harness:

# Define ~/.spindle/profiles/my-endpoint/profile.json, then:
spool_id = spin("Summarize this module", harness="my-endpoint", working_dir="/proj")
spool_id = spin("Quick pass", harness="my-endpoint", model="fast")  # profile model_aliases
result = unspool(spool_id)

Built-in harness names (claude-code, codex, gemini, kimi) always win over a same-named profile. spin_harnesses() lists lodged profiles alongside the built-ins.

profile.json fields

All fields are optional except that the file must parse as a JSON object:

• description — one-liner shown in spin_harnesses()
• harness — base harness (default "claude-code"; only "claude-code" is supported as a base in v1)
• model — default model (a caller-passed model still wins)
• model_aliases — profile-scoped alias map applied to the caller's model
• base_url — sets ANTHROPIC_BASE_URL
• api_key — sets ANTHROPIC_API_KEY
• config_dir — sets CLAUDE_CONFIG_DIR (defaults to an isolated per-profile dir when base_url is set, so the child doesn't load your real ~/.claude)
• env — arbitrary extra child env vars
• extra_args — flags appended verbatim to the claude CLI

Secret resolution

Every string value is resolved fresh at spawn time (so rotated secrets take effect on respin):

1. ${ENV_VAR} is expanded from the environment; an unset var is left literal and a warning is logged.
2. A value containing op:// is resolved via strongbox inject (if strongbox is on PATH) or op inject (if op is on PATH); if neither exists it's left literal. This keeps 1Password/strongbox an optional convenience — the ${ENV} path needs no external tool.

See examples/profiles/anthropic-compatible/ for a worked example and the full schema reference.

API

Unified API (works with all harnesses)

Tool	Purpose
spin(prompt, permission?, shard?, system_prompt?, working_dir?, allowed_tools?, tags?, model?, timeout?, harness?)	Spawn agent, return spool_id
unspool(spool_id, full?, offset?, limit?)	Get result (auto-detects harness, non-blocking; truncates huge results to head+tail by default)
respin(session_id, prompt)	Continue session (auto-detects harness)

The MCP tool spindle_repair(spool_id, attest_dead) provides the same human-attested indeterminate settlement.

Both repair entry points refuse unless the owner episode is lock_bound or accepted, its exact lock inode is released, and the recorded owner and watchdog are affirmatively dead. Settlement also refuses while the recorded provider process is observably alive, because it releases the spool's capacity and a running provider means the work is not over. The attestation covers the owner, watchdog, and provider; it claims neither cleanup nor a provider outcome, and the retained spool records an indeterminate abandonment.

spin() parameters:

• prompt (required): The task for the agent
• harness (optional): "claude-code" (default), "codex", "gemini", or "kimi"
• working_dir (optional for Claude, required for Codex/Gemini/Kimi): Project directory
• permission (optional): "readonly" (alias "manual"), "careful" (default, = auto), "full", "shard", "careful+shard", "research", "research+shard", "auto", "auto+shard" (readonly/manual cannot be combined with shard intent; auto variants are Claude-only; other names map to harness-specific enforcement, and Kimi accepts them for compatibility but has no "careful" approval mode)
• model (optional): Model to use ("sonnet", "opus", "fable", "haiku" rolling aliases or pinned "fable-5.1", "opus-5.5", "sonnet-5", "opus-5" for Claude; "astra", "sol", "terra", "luna", "reserve" for Codex; "flash", "pro", "video" for Gemini; "k3", "latest", "thinking", "k2.7-code", "k2.6" for Kimi)
• timeout (optional): Auto-kill after N seconds
• tags (optional): Comma-separated tags for organization
• shard (optional): Create isolated git worktree (can also use permission="shard")
• system_prompt (optional): Custom system prompt for Claude Code
• allowed_tools (optional): Override permission profile with explicit tool list

Spool Management (works with all harnesses)

Tool	Purpose
spools()	List all spools
spin_wait(spool_ids?, mode?, timeout?, time?)	Block until spools complete, or wait for duration
spin_sleep(duration)	Sleep for a duration (90m, 2h, 30s, HH:MM)
spin_drop(spool_id)	Cancel by killing process
spool_search(query, field?)	Search prompts/results
spool_results(status?, since?, limit?)	Bulk fetch with filters
spool_grep(pattern, spool_id?, context?)	Regex search results; pass spool_id for line-level matches with context in one result
spool_retry(spool_id)	Re-run with same params
spool_peek(spool_id, lines?)	See partial output while running
spool_dashboard()	Overview of running/complete/needs-attention
spool_stats()	Get summary statistics
spin_harnesses()	List available harnesses, models, and defaults
spool_export(spool_ids, format?, output_path?)	Export to file
shard_status(spool_id)	Check shard worktree status
shard_merge(spool_id, keep_branch?)	Merge shard to master
shard_abandon(spool_id, keep_branch?)	Discard shard

Storage

Spools persist to ~/.spindle/spools/{spool_id}.json:

{
  "id": "abc12345",
  "status": "complete",
  "prompt": "...",
  "result": "...",
  "session_id": "...",
  "permission": "careful",
  "allowed_tools": "...",
  "tags": ["batch-1"],
  "shard": {
    "worktree_path": "/path/to/worktrees/abc12345-...",
    "branch_name": "shard-abc12345-...",
    "shard_id": "..."
  },
  "pid": 12345,
  "created_at": "2025-11-26T...",
  "completed_at": "2025-11-26T..."
}

Service commands

spindle install-service  # Install background service (Linux/macOS)
spindle start            # Start via systemd (or background if no service)
spindle reload           # Drain (wait for spools to finish), then restart
spindle reload --force   # Restart immediately, interrupting in-flight spools
spindle status           # Health of the service on this port
spindle doctor           # Diagnose this install (see Install, above)
spindle serve --http     # Run MCP server directly

Background Service

For persistent background operation:

# Install and enable the service (Linux or macOS)
spindle install-service

# Start it
spindle start

Linux: Writes a systemd user service to ~/.config/systemd/user/spindle.service

macOS: Writes a launchd plist to ~/Library/LaunchAgents/com.spindle.server.plist and loads it immediately

The generated unit bakes in the current PATH, SPINDLE_PORT, and SPINDLE_HOME. PATH matters: a systemd user unit otherwise starts with a minimal one and cannot find claude/codex/gemini, which shows up much later as a spool that fails at spawn.

Because that PATH is a snapshot, it goes stale — installing a harness somewhere new, or a node upgrade relocating codex, leaves the service unable to find what your shell finds fine. spindle doctor compares the two and says so, naming the service. The fix is to re-run install-service --name <name> --port <port> --force from a shell with the right PATH, then spindle reload --name <name>.

--force overwrites a service file spindle wrote earlier. If the file does not carry spindle's marker — because you wrote it, or because you copied an example and edited it — --force first copies it aside as <name>.service.bak-<date> and says where. Ownership of a service file cannot be inferred reliably, so spindle keeps a copy rather than guess.

Then spindle reload restarts the service to pick up code changes.

Two installs on one machine

A released wheel and a working checkout can serve at once, as long as the second gets its own service name, port, and spool store:

SPINDLE_HOME=~/.spindle-release spindle install-service --name spindle-release --port 8042
spindle start --name spindle-release
spindle doctor --port 8042

spindle status and spindle doctor compare the version and package path the service reports at /health against the CLI that is asking. A service that is a different install, or the same install running older code, is reported as such rather than counted as healthy — so a fresh install cannot mistake an existing service for its own. Point either command at a specific service with --port, or set SPINDLE_PORT once in the environment.

Windows

On Windows, run spindle manually:

spindle serve --http

Or use NSSM to create a Windows service.

WSL

In WSL2 with systemd enabled, spindle install-service works like native Linux. If systemd isn't enabled, you'll get instructions to enable it or run manually.

Hot Reload (MCP tool)

From within Claude Code, call spindle_reload() to pick up code changes. By default it drains first: it returns immediately and restarts in the background once no spools are running or pending, so in-flight agents finish cleanly. New spins are still accepted while draining; the restart happens at the next idle moment. Pass force=True to restart immediately (the old behavior), which may interrupt in-flight spools and leave them to orphan recovery on the next boot. If both the logical owner and its containment watchdog died before either could publish cleanup proof, an unforced reload refuses with the affected spool ID instead of waiting forever or inventing a terminal result. --force/force=True still bypasses that refusal and may interrupt other live spools.

Configuration

Environment variables:

Variable	Default	Description
SPINDLE_HOME	~/.spindle	Spool store ($SPINDLE_HOME/spools) and lodged profiles
SPINDLE_PORT	8002	Port used by serve --http, status, and doctor
SPINDLE_HOST	127.0.0.1	Host used by the same
SPINDLE_MAX_CONCURRENT	15	Maximum concurrent spools
SPINDLE_CLAUDE_STREAM_DRIVER	enabled	Claude stream driver; set to 0, false, no, or off for the legacy one-shot rollback path
SPINDLE_UNSPOOL_MAX_CHARS	50000	Results longer than this are truncated to head+tail by unspool()
SPINDLE_UNSPOOL_HEAD_CHARS	12000	Chars kept from the start of a truncated result
SPINDLE_UNSPOOL_TAIL_CHARS	12000	Chars kept from the end of a truncated result

Storage location: ~/.spindle/spools/, or $SPINDLE_HOME/spools/ when set. The spool store must be on a persistent local filesystem. Network and distributed filesystems (such as NFS, SMB/CIFS, and CephFS), network-backed FUSE or object-storage mounts, VM shared folders, container writable overlay layers, and ephemeral filesystems such as tmpfs are unsupported because they may not preserve Spindle's locking, inode-identity, and crash-durability assumptions. In a container, bind-mount a directory from a local host filesystem; in WSL, keep the store in the Linux filesystem rather than under /mnt/c.

spindle doctor reports the store it resolved, whether it is writable, and whether its existing ownership artifacts are healthy. It does not verify the filesystem semantics above.

How It Works

1. spin() spawns a detached CLI process (claude, codex, gemini, or kimi-cli) with the given prompt
2. The process runs in background, writing output to temporary files
3. A monitor thread polls for completion
4. unspool() returns the result once complete (non-blocking check)
5. Spool metadata persists to JSON files, surviving server restarts

For shards:

1. A git worktree is created with a new branch
2. The agent runs inside that worktree
3. After completion, merge back with shard_merge() or discard with shard_abandon()

Limits

• Max 15 concurrent spools (configurable via SPINDLE_MAX_CONCURRENT)
• 24h auto-cleanup of old spools
• Orphaned spools (dead process) marked as error on restart

Contributing

See CONTRIBUTING.md for development setup and guidelines.

License

MIT - see LICENSE.