Private, fast inference for coding agents

Open-weight models. Zero data retention. Zero training. Long-context, multi-region speed.

zro.moonmath.ai · Pricing · Integrations

zro is the CLI that connects your coding agent to the Zro inference endpoint. It launches the agent of your choice with isolated, Zro-owned configuration — so it never touches your everyday agent setup, credentials, or session history.

No training on your prompts or completions. Zero request retention. Runs on privacy-forward, multi-region infrastructure. Optimized for the long-context, tool-heavy workloads coding agents depend on.

npm install --global @moonmath-ai/zro
zro login
zro claude

Run zro models for the models available to your key, or see zro.moonmath.ai/pricing. Region availability is shown when you create an API key.

---

Quick start

zro login                 # website or API key login
zro claude                # open Claude Code on the default model
zro codex -m <model>      # pick a model for this session
zro models                # browse the model catalog
zro status                # connection, installed tools, last session, usage

zro <tool> --install installs a missing agent and opens it in one step.

---

Installation

• Requirements: Node.js 22 or later on macOS, Linux, or Windows.
• Install: npm install --global @moonmath-ai/zro (also pnpm add --global @moonmath-ai/zro).
• Upgrade: zro install --upgrade (upgrade zro); zro install <tool> --upgrade or zro install claude@2.1.105 / zro install claude --version 2.1.105 for an agent.
• Agent-assisted setup: SETUP_PROMPT.md has copy-paste prompts for end-user installs and local development checkouts.

Authentication

Sign in once, then launch any supported agent:

zro login                    # website or API key login
zro login --manual           # enter an API key directly
zro login --no-browser       # print an approval URL (e.g. remote box)
zro logout                   # remove the stored login

• Website login creates a revocable API key after approval, encrypted to an ephemeral CLI-held public key. Set ZRO_DEVICE_NAME to label the device.
• ZRO_API_KEY takes precedence over the stored key at ~/.config/zro/credentials.json; Codex App keeps its own copy under ~/.config/zro/codex-app/.env; zro logout removes both.
• The credential is verified against the inference API before any launch.

Alternate inference URL and no-auth mode

Point inference at another OpenAI-compatible gateway with --inference-url or ZRO_INFERENCE_URL. A root URL or one ending in /v1 both work; trailing slashes are normalized.

zro claude --inference-url https://gateway.example
ZRO_INFERENCE_URL=https://gateway.example zro codex

Precedence is --inference-url > ZRO_INFERENCE_URL > ZRO_ENDPOINT_ROOT > https://zro.moonmath.ai. The override only changes inference traffic; login, the authenticated catalog, feedback, and MCP stay on ZRO_AUTH_URL / ZRO_ENDPOINT_ROOT.

For an inference endpoint with no authentication and no Zro account, add --no-auth or set ZRO_NO_AUTH=1. No-auth mode requires an explicit --inference-url or ZRO_INFERENCE_URL.

zro claude --no-auth --inference-url https://gateway.example
zro models --no-auth --inference-url https://gateway.example
ZRO_NO_AUTH=1 ZRO_INFERENCE_URL=https://gateway.example zro status

In no-auth mode the CLI skips login and API-key checks, sends no Authorization headers, and uses a fixed non-secret placeholder key only where an agent requires a non-empty value. Public model metadata comes from the main webapp (/api/cli/models/public) and is intersected with the gateway's live /v1/models list, so zro models only shows models the gateway can serve. That metadata is cached separately at ~/.cache/zro/model-catalog-public.json. The Zro MCP server is omitted from every agent config because it requires an account, and zro feedback is unavailable in this mode.

---

Supported coding agents

Agent	Install	Notes
Claude Code	zro install claude	Anthropic's coding agent
Codex CLI	zro install codex	OpenAI's terminal agent
Codex App	zro install codex-app	OpenAI's desktop app
Kilo Code	zro install kilo	Kilo's terminal coding agent
Oh My Pi	zro install omp	Power-user terminal agent
OpenCode	zro install opencode	Open-source coding agent
OpenCode 2	zro install opencode2	OpenCode 2.0 client/server agent
Grok Build	zro install grok	xAI's coding agent
Hermes	zro install hermes	Nous Research's agent
OpenClaw	zro install openclaw	Personal AI assistant
Pi	zro install pi	Minimal coding agent
Prime Agent	zro install prime	Prime Intellect's self-improving RLM agent

Each adapter launches the installed tool as a child process with isolated, Zro-owned configuration. Session files are temporary; Codex App uses a persistent Zro-owned home. Pass native tool arguments after -- (for example zro oc -- --help). The last tool/model pair is remembered and reopened with zro again.

---

Usage

zro                         # interactive agent picker
zro claude                  # open directly on the default model
zro codex -m <model>        # choose a model for this session
zro omp -m <model>          # launch Oh My Pi through Zro
zro kilo -m <model>         # launch Kilo Code through Zro
zro oc -- --help            # short aliases + native tool arguments
zro again                   # reopen the last tool/model pair
zro models                  # readable model catalog
zro install claude@2.1.105  # pin an agent version
zro feedback                # send feedback to the Zro team
zro claude --dry-run        # secret-safe session preview
zro codex --json            # machine-readable preview
zro claude --no-auth --inference-url https://gateway.example
                            # launch against a no-auth inference endpoint

---

Pricing

Plans are spend-based with expected usage, so actual totals vary by model mix, prompt shape, output length, and cache reads. See the live pricing page.

Plan	Per month	Yearly	Expected usage
Pro	$20	$16/mo	~1B tokens
Max	$60	$48/mo	~5B tokens

• Usage packs are available with or without a subscription and expire after 90 days.
• Monthly spend resets on the same calendar cadence for monthly and yearly subscribers.
• There are no prorated refunds on cancellation; access stays active to the end of the period.

---

Safety

• Normal agent configs are never edited.
• Temporary session files live under ~/.cache/zro/sessions and are removed when the agent exits.
• Kilo Code and Oh My Pi run with a temporary home and XDG profile; only sanitized preferences and safe assets are copied in. Provider credentials, MCP definitions, sessions, databases, dependency trees, and secret files stay outside the profile.
• Kilo and Oh My Pi telemetry, OTLP export, automatic updates, sharing, and remote control are disabled; model and MCP requests still go to the configured Zro endpoints.
• API keys are masked in human and JSON previews. --dry-run writes nothing and starts nothing.
• macOS, Linux, and Windows are supported.

See SECURITY.md for the security model and private vulnerability reporting.

---

Contributing

Zro is a small, dependency-light TypeScript CLI (runtime deps: json5 and yaml). Bug reports, feature requests, and pull requests are welcome. See CONTRIBUTING.md.

---

License

Zro is licensed under the MIT License.