Claw Orchestrator CLI
One runtime for Claude Code, Codex, OpenCode and more: a run counts as done only when checks the agent did not write pass
- 601 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- The runtime launches coding agents that edit files and run commands in the workspace
- Contract checks run commands and HTTP requests on your behalf
- Token spend goes against your own subscriptions and keys with several vendors, and the spend ceiling is set with the maxBudgetUsd parameter
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add cli/claw-orchestratorDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- The runtime launches coding agents that edit files and run commands in the workspace
- Contract checks run commands and HTTP requests on your behalf
- Token spend goes against your own subscriptions and keys with several vendors, and the spend ceiling is set with the maxBudgetUsd parameter
This is third-party code. Review the repository files before installing.
What it does
Claw Orchestrator runs Claude Code, Codex, Antigravity, Grok Build, OpenCode and any custom CLI as persistent sessions behind one API. The agent does not pass the verdict on its own work: you declare the checks, the runtime executes them and reads the exit codes, and the evidence stays on disk. Checks can be commands, HTTP probes, screenshots, diff policy and file assertions. A run with a contract ends as verified, refuted or unverified. If the agent edits the tests that the contract runs, the run is refuted. The clawo solve command sets the task and checks, clawo verify shows each check and why it failed, and clawo runs prints a ledger of turns across all engines with model, tokens, cost and verdict. The runtime enforces the maxBudgetUsd cap on any engine. The clawo fanout command asks several engines the same question at once, and session_handoff moves a live conversation to another engine in the same workspace. There are also councils of agents in separate git worktrees that vote, a planner-coder-reviewer loop and durable workflows with checkpoints. The runtime can be called as a CLI, as the clawo-mcp MCP server with 78 tools (the list can be narrowed with the CLAWO_MCP_TOOLS variable), as an ACP agent for Zed and JetBrains with clawo acp, through the OpenAI-compatible /v1/chat/completions endpoint, from TypeScript and as an OpenClaw plugin. The local server and dashboard start with clawo serve.
Who it is for. For developers who work with several coding agents and do not accept a result on the agent's word: they want a shared launcher, a spend ledger and independent checks.
Good fit when
- An agent writes all tests pass and the runtime itself must confirm it: clawo solve with the --check flag runs the command and reads the exit code
- One question must go to Claude Code, Codex and Grok at once with a combined answer: clawo fanout with --synthesize handles it
- A stuck Claude session has to continue in Codex with the same conversation: session_handoff moves it into the same workspace
- Spend per engine must be visible and capped: clawo runs shows the turns and maxBudgetUsd sets a ceiling
- You need to call agents from Zed or JetBrains or from a client that speaks the OpenAI format
Not a fit when
- You need an interface for parallel worktree sessions: Claw Orchestrator is a headless runtime, and interfaces such as claude-squad, Conductor and Vibe Kanban cover that
- You cannot define any checks: a run without a contract ends as unverified and the result is never confirmed
- Your work stays inside one vendor, only Claude Code subagents for example: the runtime is overkill there
Example request
Fix the failing test in price.js with Codex and accept the work only if npm test and the independent price.spec.mjs check passLimitations
Node.js 22 or newer and at least one installed and logged-in engine CLI are required: claude, codex, agy, grok or opencode. The server writes the local access token to ~/.openclaw/server-token, and the browser signs in through a link carrying that token.
In the demonstration run two launches ended with an error: Grok Build hit its free-tier usage limit. The author notes that the run covers a single task and is no benchmark. The table of compatible engine versions is refreshed by the weekly scripts/sweep.ts run, so between runs your engine version may be newer than the tested one.
How to disable. Stop the clawo serve process and remove the package with npm uninstall -g @enderfga/claw-orchestrator. If you installed the OpenClaw plugin with install.sh, remove its entry from ~/.openclaw/openclaw.json. In Claude Code the MCP block is removed with claude mcp remove clawo.
FAQ
How do I connect Claw Orchestrator to Claude Code?
Install the package with npm install -g @enderfga/claw-orchestrator, which puts clawo-mcp on the PATH, and run claude mcp add clawo -- clawo-mcp. Then ask the agent to call workflow_start with the solve template, a task and a contract such as an npm test check.
How does the runtime decide a task is done?
It runs the checks you declared itself: commands, HTTP probes, screenshots, diff policy and file assertions. A run with a contract cannot reach completed until all checks pass, and the evidence for each check is saved on disk.
Which agents does Claw Orchestrator work with?
Claude Code, Codex, Antigravity (agy), Grok Build and OpenCode, plus any CLI that runs as a subprocess and is wired up as a custom engine. As an MCP server it connects to Claude Desktop, Cursor, Cline, Continue, Zed, Windsurf, Goose and Hermes Agent.
How does the runtime differ from claude-squad or Conductor?
Those are interfaces that run agents side by side in worktrees. Claude Code subagents and Codex's parallel attempts stay within one vendor. Claw Orchestrator is a headless runtime: several vendors behind one API, durable workflows and a verification step the agents cannot edit. The tools can be used together.
How do I cap spend on agents?
Set maxBudgetUsd: the runtime enforces the ceiling on every engine, including Codex, Grok, agy and OpenCode. The clawo runs --since 1h ledger shows turns with model, tokens and cost.
Discussion
No questions or reviews yet.
Related
CLI toolsMulti-agent orchestration for Claude Code: teams, roles, autopilot and task routing across skills and subagents
CLI Agent Orchestrator
CLI Agent Orchestrator (CAO)
A CLI that coordinates several coding agents: a supervisor delegates tasks to specialists in parallel or in sequence across isolated tmux sessions
A compact OpenClaw alternative: a personal assistant on Claude Agent SDK with agents in isolated containers and messengers added via skills
Self-hosted system for a team of AI agents: org chart, tasks, budgets and oversight of Claude Code, Codex and OpenClaw
Workflows for Claude Code: /do, /omo, /bmad-pilot and more commands, with Codex, Gemini or OpenCode writing the code through codeagent-wrapper
A CLI orchestrator that loops Claude Code, Codex, Gemini CLI and other agents until a task is done and blocks incomplete work with tests and lint
Comparison with similar tools
Entries that share the main search query, and how they differ.
| Entry | Official | Repository stars | Main difference |
|---|---|---|---|
| Claw Orchestrator (this entry) | no | 601 | |
| CLI Agent Orchestrator | no | 1,489 | a supervisor hands tasks to specialists in separate tmux sessions |
| Babysitter | no | 1,880 | the workflow is described in code and decisions go into an immutable journal |
| comanda | no | 331 | orchestrator for Claude Code |
README in short
The description opens with a slogan and an animation of a single experiment: one bug, one acceptance contract, five engines, three runs each. A table maps user goals to runtime capabilities, followed by a quick start, the full feature list with links to reference files, integrations (CLI, OpenClaw plugin, MCP, ACP), a table of compatible engine versions and an account of how a weekly sweep keeps it current.