Claw Orchestrator

A runtime for coding agents. Wrap Claude Code, Codex, Antigravity, Grok Build, OpenCode, or any custom CLI as persistent programmable sessions; coordinate them in multi-agent councils; run autonomous Planner / Coder / Reviewer loops; or hand a five-question interview to an Opus council that ships a deployed web app at localhost:19000/forge/<slug>/.

Coding CLIs are designed for humans at terminals. Claw Orchestrator turns them into headless engines and stacks an agent platform on top: a 78-tool API that scales from a single session call up to a fully generated, deployed web app — reachable through the CLI, the OpenClaw gateway, the Model Context Protocol, or directly from TypeScript, and visible through an embedded three-tab dashboard.

https://github.com/user-attachments/assets/fbd2b0ea-28d8-4387-9894-c29cf15ba030

Features

| Capability | What it does | Reference | | --------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------- | | Persistent Sessions | Long-lived coding agents kept alive across requests, with full context, tool, model, and worktree control. | sessions.md | | Multi-Engine Runtime | One interface over Claude Code, Codex, Antigravity (agy), Grok Build, OpenCode, and arbitrary custom CLIs. | multi-engine.md | | Session Handoff | Move a live conversation to another engine or model — a stuck Claude session into Codex, an expensive model into a cheaper one. The new session picks up where the old one stopped, in the same workspace; the old one keeps running. | sessions.md | | Multi-Agent Council | Parallel agents in isolated git worktrees, voting on consensus until they agree. | council.md | | Fan-out | Run one task across N engine/model agents in parallel and collect their answers, with an optional synthesis pass — the cross-engine best-of-N / diverse-perspective primitive (no rounds or worktrees). | tools.md | | ultracode | session_start({ ultracode: true }) lets Claude orchestrate a dynamic JS workflow and fan out to subagents per task (Claude engine). | tools.md | | Autoloop | Three-agent autonomous workspace iteration with independent engine/model selection for Planner, Coder, and Reviewer. Chat with the Planner; it spawns Coder + Reviewer into a self-iterating subloop and pushes you on regression, target-hit, or decision points. | autoloop.md | | Ultraapp | A three-agent Opus council turns a five-question interview into a deployed web app — Tailwind UI, BYOK, file-queue runtime, smoke test, all live at localhost:19000/forge/<slug>/. | ultraapp.md | | Embedded Dashboard | Three-tab UI for Autoloop, Council, and Forge with sidebar lifecycle controls, per-run live event streaming, and cookie-based auth via a /login redirect. | dashboard.md | | OpenAI-Compatible Proxy | POST /v1/chat/completions translates OpenAI requests into native Anthropic, OpenAI, and Google calls and streams responses back in OpenAI shape. Point any OpenAI-SDK client at the orchestrator without changing call sites. | openai-compat.md | | Durable Run Kernel | Declarative workflows over agent / fanout / council / verifier / human_gate / router / subflow / autoloop / ultraapp_* nodes. Every state transition is checkpointed, so a run survives a process restart and resumes at the node boundary. Retry, per-node timeout, cancel, steer, and bounded loops come from the kernel instead of from five hand-rolled state machines. | workflow.md | | Verification Plane | Acceptance contracts the runtime executes itself — commands, HTTP probes, screenshots, diff policy, file assertions — producing an evidence bundle on disk. A run carrying a contract cannot reach completed unless it passes, and one without a contract completes as unverified rather than claiming success. The tests a contract runs are held to what the run started with, so a run cannot pass by editing them. | verification.md | | Run Ledger & Spend Caps | Every turn on every engine is appended to a durable JSONL ledger — engine, model, tokens, cost, duration, and the council/fanout/autoloop it belonged to — queryable with clawo runs after a restart. Rows carry both the engine's self-report (ok) and the runtime's own measurement (verified), kept apart. maxBudgetUsd is enforced by the runtime, so a cap holds on Codex, Grok, agy and OpenCode too, not just Claude Code. | observability.md |

The full 78-tool surface is enumerated in tools.md.

Quick Start

npm install -g @enderfga/claw-orchestrator
clawo serve   # dashboard at http://127.0.0.1:18796/dash

import { SessionManager } from '@enderfga/claw-orchestrator';

const manager = new SessionManager();
await manager.startSession({ name: 'fix-tests', engine: 'claude', cwd: '/project' });
const result = await manager.sendMessage('fix-tests', 'Fix the failing tests');

Integrations

Standalone CLI

clawo serve                                            # dashboard + HTTP server on :18796
clawo session-start fix-tests --engine claude --cwd .  # start a session
clawo session-send fix-tests "Fix the failing tests"   # send into it

Every command is documented in cli.md.

OpenClaw Plugin

curl -fsSL https://raw.githubusercontent.com/Enderfga/claw-orchestrator/main/install.sh | bash

Installs via npm, registers the plugin in ~/.openclaw/openclaw.json, restarts the gateway. All 78 tools become available to every OpenClaw agent.

Model Context Protocol Server

npm install -g @enderfga/claw-orchestrator   # clawo-mcp is now on PATH

Register clawo-mcp with any MCP-compatible host: Hermes Agent, Claude Desktop, Cursor, Cline, Continue, Zed, Windsurf, Goose, and others. Per-host stdio-config snippets and the CLAWO_MCP_TOOLS allowlist for tight tool budgets are in mcp.md.

Agent Client Protocol Agent

clawo acp        # or the dedicated binary: clawo-acp

MCP gives tools to an agent; ACP makes you be the agent. clawo acp speaks Agent Client Protocol over stdio, so Zed, JetBrains, Neovim, Emacs, the VS Code ACP extension — or dsh via its subagent-acp provider — can drive Claw Orchestrator as their coding agent.

Every other agent in that ecosystem is a single agent. This one is a fleet: the model selector is grouped by engine, so one dropdown holds Claude, Codex and Grok models at once and switching it switches engine mid-session, and /council, /ultraplan and /ultrareview run multi-agent orchestrations from the chat box. Setup, the dsh YAML block, and the cancellation and permission limitations are in acp.md.

Engine Compatibility

| Engine | CLI | Tested Version | | ----------- | ---------- | -------------- | | Claude Code | claude | 2.1.274 | | Codex | codex | 0.154.0 | | Antigravity | agy | 1.2.5 | | Grok Build | grok | 1.0.34 | | OpenCode | opencode | 1.18.31 | | Custom CLI | any | — |

Any coding CLI that runs as a subprocess can be wired up as a custom engine — see multi-engine.md.

How the engine table stays honest

The versions above are not typed in — they are what the weekly sweep last ran. scripts/sweep.ts measures each core engine's installed, pinned and upstream version, diffs the flags the wrapper passes against the binary's --help, runs one live turn through the real wrapper class, and smokes the ACP and MCP entry points. It has no LLM in it, so the thing that reports a wrapper as broken cannot share the wrapper's failure modes.

scripts/sweep-workflow.json wraps it as a durable run on this project's own kernel: verifier → router → an agent that drafts the alignment on a sweep/<date> branch → a human gate. That is the bounded form of recursive self-improvement this project practises — the loop measures, proposes and verifies; a person merges. The orchestrator never edits its own code unattended, on purpose: the recovery path has to stay simpler than what it recovers. Its first scripted run found a default model an engine had silently dropped, which three weeks of by-hand sweeps had walked past.

Contributing

See CONTRIBUTING.md. Run npm run build && npm run lint && npm run format:check && npm run test before submitting.

License

MIT — see LICENSE.