Claw Orchestrator
Turns Claude Code, Codex, Antigravity and other coding CLIs into programmable, persistent engines you can orchestrate in sessions, councils and autonomous loops.
- Source repo
- Enderfga/claw-orchestrator
- Stars
- ★ 603
- Last updated
- today
- License
- MIT
- Primary language
- TypeScript
- FA score
- 61/100 · Some gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude Code · OpenAI API · Claude API
- Cost
- Free software; you pay for model usage
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- Developers who need to drive Claude Code or Codex headlessly from scripts or backend services with full context retained across requests.
- Not a fit if
- Teams unwilling to install and maintain the underlying coding CLIs
- Users sensitive to multi-agent token costs without budget caps
- Those needing production hosting; Ultraapp deploys only locally
- Source review
- 61/100 · Some gaps
What does this agent do, and when should you use it?
Claw Orchestrator is a runtime for coding agents that wraps human-facing coding CLIs — Claude Code, Codex, Antigravity, Grok Build, OpenCode, or any custom CLI — as persistent, programmable sessions. It exposes a 77-tool API that scales from a single session call to a fully generated, deployed web app, reachable via the CLI, the OpenClaw gateway, an MCP server, an ACP agent, or TypeScript, with a three-tab embedded dashboard for visibility. Core primitives include multi-agent councils (parallel agents in isolated git worktrees voting to consensus), a Planner/Coder/Reviewer Autoloop, cross-engine fan-out, and Ultraapp, which turns a five-question interview into a deployed web app. A durable run kernel organizes declarative workflows over agent/fanout/council/verifier/human_gate nodes with checkpointed state transitions that survive process restarts. Every turn is appended to a durable JSONL run ledger recording engine, model, tokens, cost and duration, and the runtime enforces maxBudgetUsd spend caps across all engines.
After installing, clawo serve starts a dashboard and HTTP server at 127.0.0.1:18796/dash. You create named persistent sessions via the SessionManager class or CLI (clawo session-start / session-send), specifying engine, model, cwd and worktree. The council capability runs parallel agents in isolated git worktrees and votes until consensus; fanout distributes one task across N engine/model agents with an optional synthesis pass; ultracode lets Claude orchestrate a dynamic JS workflow fanning out to subagents per task. Autoloop runs Planner, Coder and Reviewer as a self-iterating subloop; Ultraapp uses a three-agent Opus council to turn a five-question interview into a Tailwind web app deployed at localhost:19000/forge/<slug>/. The verification plane executes acceptance contracts (commands, HTTP probes, screenshots, diff policy, file assertions) producing an on-disk evidence bundle — runs with a contract cannot reach completed unless they pass, and runs without one complete as unverified. POST /v1/chat/completions provides an OpenAI-compatible proxy; clawo-mcp exposes an MCP server; clawo acp lets editors drive it as their coding agent over Agent Client Protocol.
- Developers who need to drive Claude Code or Codex headlessly from scripts or backend services with full context retained across requests.
- Engineers wanting cross-model validation of a single solution: use council or fanout to run multiple engines in parallel in isolated worktrees and take consensus or best-of-N.
- Teams who want an autonomous loop continuously iterating a codebase: chat with the Planner while Coder and Reviewer loop, surfacing regressions, target hits and decision points.
- Users integrating coding agents into MCP hosts (Claude Desktop, Cursor, Zed, Cline) or driving it from Zed, JetBrains, Neovim or Emacs via ACP.
- Projects on the OpenAI SDK that want to transparently switch underlying Anthropic/OpenAI/Google models by pointing at /v1/chat/completions without changing call sites.
- Teams needing auditable spend: query per-turn tokens, cost and attribution with clawo runs and set hard maxBudgetUsd caps.
How do you install or deploy this agent?
Prerequisites: a Node.js/npm environment and at least one supported coding CLI (claude, codex, agy, grok, opencode) or any custom CLI runnable as a subprocess. Install globally: npm install -g @enderfga/claw-orchestrator. This puts clawo-mcp on PATH. The OpenClaw plugin installs via curl -fsSL https://raw.githubusercontent.com/Enderfga/claw-orchestrator/main/install.sh | bash, which registers the plugin in ~/.openclaw/openclaw.json and restarts the gateway.
How do you use this agent?
Start the server: clawo serve, dashboard at http://127.0.0.1:18796/dash. CLI sessions: clawo session-start fix-tests --engine claude --cwd . then clawo session-send fix-tests "Fix the failing tests". TypeScript: import { SessionManager } from '@enderfga/claw-orchestrator'; await manager.startSession({ name: 'fix-tests', engine: 'claude', cwd: '/project' }); then manager.sendMessage('fix-tests', '...'). Other entry points: clawo acp (or clawo-acp) to be driven by editors over ACP; register clawo-mcp with any MCP host; OpenAI-compatible endpoint at POST /v1/chat/completions.
What are this agent's strengths and limitations?
- A genuine multi-engine runtime: one interface over Claude Code, Codex, Antigravity, Grok Build, OpenCode and arbitrary custom CLIs, with mid-session engine switching.
- Durable run kernel with checkpointing: declarative workflows over verifier, human_gate, router and other nodes resume at node boundaries after process restarts, with retry, timeout, cancel and steer built in.
- Runtime-enforced verification and spend caps: contract-carrying runs cannot reach completed without passing, and maxBudgetUsd holds on Codex, Grok, agy and OpenCode too, not just Claude Code.
- Four integration surfaces (CLI, OpenClaw plugin, MCP, ACP) cover terminal-to-editor call paths, backed by a three-tab live dashboard.
- Hard dependency on underlying coding CLIs: you must install and maintain claude, codex, agy, grok or opencode, with tested version requirements (e.g. claude 2.1.260, codex 0.153.2).
- Cost scales with the multi-agent design: council, fanout and Autoloop invoke multiple engines concurrently, which can be expensive without budget configuration.
- Ultraapp's deployment boundary is local (localhost:19000/forge/<slug>/), aimed at local demos rather than production hosting.
- The editor ACP integration has documented cancellation and permission limitations; read acp.md to understand the boundaries before adopting.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Claw Orchestrator This agent | 61 · Some gaps | CLIFree + model costs | ★ 603 | today | TypeScript | Codex · Claude Code · OpenAI API · Claude API |
| TAKT Workflow Orchestrator | 71 · Some gaps | CLIFree + model costs | ★ 1.4k | today | TypeScript | Codex · Claude Code · OpenAI API · Claude API |
| Lody | 63 · Some gaps | Self-hosted service | ★ 1.7k | today | TypeScript | Codex · Claude Code |
| fx Coding Agent | 65 · Some gaps | CLIFree + model costs | ★ 3.8k | today | Zig | ChatGPT · Codex · OpenAI API |
How does FollowAgents rate this agent?
Why each dimension lost points
The source centralizes CLI spawning in engine-spawn.ts, while the README identifies the local server, token location, acceptance checks, and spend ledger; workflow support includes cancellation, resume, and rollback. These give some evidence for data flow and recovery. Deductions: the candidate supports arbitrary custom CLIs, MCP tools, and generated deployments, but the sample does not adequately show permission scoping, confirmation for external effects, or sensitive-data protections. Server-environment secret resolution is mentioned without showing storage or leak safeguards. Repository authorship and a publishing path are traceable, but publisher identity is unverified.
The workflow nodes, templates, and README are broadly consistent about runtime verification, persistence, and recovery; CI specifies build, lint, formatting, and test steps. Deductions: CLI and multiple external engines remain runtime dependencies, with no availability safeguards shown in this sample. Timeouts, retries, and status handling are described, but evidence for user-facing error messages is limited. This static review does not claim that tests ran.
The README, skill description, and package metadata cover terminal, scripting, MCP, ACP, SDK, OpenClaw plugin, and multi-engine scenarios, with concrete users and workflows. Deductions: the skill's trigger list is broad and may match unrelated general coding or app-building requests. The sample explains capability boundaries only partly, and chiefly specifies Node 22+ and logged-in engine CLIs as environment requirements.
The README organizes features, run modes, and references, and provides a quick start, compatibility table, and examples; the MIT license, 7.20.0 version, and changelog link are clear. AGENTS.md reads as contributor architecture guidance, not mandatory product instructions. Deductions: installation has a global npm command and prerequisites, but host-specific setup is distributed across documentation not included here. Exports still include legacy Gemini/Cursor engines that are not presented as current main options. The maintainer and update path are identifiable, but no response commitment appears in the sample.
The runtime unifies multiple CLIs behind sessions, workflows, and acceptance checks, with check results, a ledger, and budget caps as useful outputs. Its value for cross-engine orchestration and independent verification is clear. Deductions: the sample gives limited concrete evidence about result formats and user outcomes; multi-process, multi-engine, and deployment workflows add setup and operating costs without guidance on cost comparisons or suitable scale.
The README links core claims to verification, workflow, and tool guides; source code shows workflow templates and node registration, while package and CI files add version, scripts, and dependency information. Deductions: the sample omits the referenced detailed guides and test implementations, and CI configuration shows intended checks but not that they passed. README claims about demos, compatibility, and capabilities do not always clearly distinguish reported facts from summary claims.
- The tool launches local engine CLIs and permits arbitrary custom CLIs. Review their permissions, working directories, and external effects before use, and avoid putting unnecessary secrets in the server environment.
- The README places the local server token at ~/.openclaw/server-token; this sample does not show detailed access controls for that token or protections for custom-engine environment variables.
- CI lists test commands, but this is a static review and does not establish that tests passed or runtime results are correct.
Maintain this project? Add the FARS badge to your README
[](https://followagents.com/en/agents/claw-orchestrator)<a href="https://followagents.com/en/agents/claw-orchestrator"><img src="https://followagents.com/badge/claw-orchestrator.svg" alt="FARS review"></a>Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v7.20.1 v7.20.1 — OpenAI-compat timeout, Planner control errors, CLAUDE_BIN
Release notes excerpt
Fixes for OpenAI-compatible request timeouts, rejected Autoloop Planner controls, and CLAUDE_BIN. - OpenAI-compatible requests wait up to 30 minutes for a turn instead of 5 (OPENAI_COMPAT_TURN_TIMEOUT_MS changes it), so a gateway agent with a long reply is no longer cut off and retried. - A rejected Planner control shows in the Autoloop reply, naming the problem for a validation error, and is recorded as a planner_control_error row in decisions.jsonl. (#162, thanks @ajmtrz) - CLAUDE_BIN takes effect for clawo serve and the MCP server. (#163, thanks @wagnerglorenz) …View on GitHub ↗ -
v7.20.0 v7.20.0 — fan-out agents that run as goals
Release notes excerpt
A fan-out agent can now run as a goal, so several agents can each work a copy of a project until it is done. - objective on a fan-out agent (engine claude or codex) runs it as a goal instead of one turn, with a goal node's rules: judged, budgeted, timed out, and resumed after stalls and capacity errors. It runs beside the agents that take one turn. - Resumable — each goal runs in a child run checkpointed on the fan-out node, so a parent that resumes after a crash reattaches to it instead of starting it over. Each result carries the goal's run id and final status. …View on GitHub ↗ -
v7.19.0 v7.19.0 — a working directory per fan-out agent
Release notes excerpt
Several fan-out agents can now write code at once, each in a directory of its own. - cwd per fan-out agent — on fan-out nodes and fanout_start, absolute or relative to the fan-out's directory. Each result records the cwd it ran in, and an unsafe one fails that agent only. - An unknown model is warned about once, not on every priced update. - Gateway models are not priced as Sonnet — under CLAWO_CLAUDE_GATEWAY, a non-Claude model is priced from pricingOverrides, not from another model's rate or the CLI's own figure; without an override its cost is 0. …View on GitHub ↗
Thanks — we use this to improve the page.