CoordClaw Multi-Agent Collaboration System
Define an AI team in plain language; agents meet, challenge, arbitrate and ship on their own. One person, one company.
- Source repo
- CoordClaw/CoordClaw
- Stars
- ★ 160
- Last updated
- 2mo ago
- License
- MIT
- Primary language
- TypeScript
- FA score
- 42/100 · Major gaps
At a glance
- Works with
- Platform-specific
- You'll need
- Typical use
- A solo developer or PM wants a small shippable artifact (e.g. a web Snake game) in about an hour with an auditable review-and-arbitration trail.
- Main limitation
- Hard dependency on the OpenClaw runtime (or a qclaw-style variant) that must be initialized first, adding an external layer and version-matching risk.
- Source review
- 42/100 · Major gaps 1 safety controls not found
What does this agent do, and when should you use it?
CoordClaw is a multi-agent collaboration system layered on the OpenClaw runtime that organizes AI according to management theory. Team structure lives in natural-language Markdown (teamsoul.md for roles, team RULE.md for collaboration rules) instead of a hard-coded DAG. Work is driven by point-to-point message loops: each agent pulls only its own unread messages via chat_manager.py inbox, then closes the loop through task_start.py, task_report.py and task_done.py structured logs. A human keeps a god's-eye view on the control panel at http://localhost:18790 — reading every message, toggling read state, and impersonating any member to intervene. Context is fully reset each round, keeping only role definitions and auditable logs to suppress context contamination. It needs Node.js >= 22 and an initialized OpenClaw (or qclaw variant), and is fully validated only on Windows so far.
CoordClaw runs as a service on port 18790; two green indicators (SSE and Gateway) confirm the runtime is ready. Creating a project copies a team template's config and scripts into .data/, where team RULE.md (five standard actions T1–T5 and ten prohibited behaviors P1–P10), teamsoul.md (per-role identity) and team.json (25 runtime parameters such as max_activations, idle_confirm_ms and six supervisor checks) define behavior. With auto-coordination on, agents execute T1 to pull unread messages with chat_manager.py inbox --reader '{name}' --last 20, T2 to create a task list via task_start.py, T3 to write a structured work log via task_report.py, T4 to reply point-to-point with chat_manager.py send --from --to, and T5 to finish via task_done.py. The panel charts token consumption using local BPE estimation, and humans can intervene through POST /api/send-message and POST /api/toggle-read.
- A solo developer or PM wants a small shippable artifact (e.g. a web Snake game) in about an hour with an auditable review-and-arbitration trail.
- Researchers studying organizational, economic or management behavior in an AI society, who need to observe and intervene at will.
- Teams that want to express AI roles, hierarchy and reporting lines in Markdown rather than wiring code for each new project structure.
- Delivery workflows where every agent's work log and point-to-point message must be persisted and reviewable.
- Teams running non-frontier or China-accessible models (Kimi, MiniMax, qclaw bundled quota) who want to test multi-agent collaboration cheaply.
How do you install or deploy this agent?
Install and initialize OpenClaw or a variant first (qclaw is recommended for mainland China users): open it once so it completes environment initialization, then exit — otherwise CoordClaw cannot discover the runtime. Then clone the repo and start it:
# Linux / macOS
node start.cjs:: On Windows you can double-click start.bat or run the same command
node start.cjsOpening http://localhost:18790 for the first time launches the installer wizard: pick a language, tick the discovered OpenClaw instance, install. If your instance lives outside standard paths, copy findplatforms.json.example to findplatforms.json, fill in the directories field and rerun the wizard. Afterwards, start OpenClaw before CoordClaw on every boot. Node.js >= 22 is required; the tested setup is Windows 11 24H2 with OpenClaw v2026.4.5 / Qclaw v0.2.32, and Linux/macOS are untested.
How do you use this agent?
1) Start OpenClaw (or variant) and then the CoordClaw service; confirm the SSE and Gateway indicators are both green. 2) Sidebar project card → new project → choose a team template (a 7-person standard team ships built in) → name it and pick a path. 3) Turn on auto-coordination, pick a recipient in the message box and send the task. 4) Watch deliveries, rejections and arbitrations in the message list; to intervene use the bottom-left send button, choosing sender identity and recipient (written straight to the database). 5) Manage skills globally from the Tools card and per member via the skill icon. For a custom team, open the Team card → New Team and follow the five-stage wizard, then review teamsoul.md and team RULE.md and register. Runtime parameters live in the project's .data/team.json. Agent-side calls look like:
python .data/scripts/chat_manager.py inbox --reader 'PM' --last 20
python .data/scripts/chat_manager.py send --from 'PM' --to 'QA'What are this agent's strengths and limitations?
- Team shape is defined entirely in natural-language Markdown (teamsoul.md / team RULE.md), so new organizational structures need no code changes.
- Point-to-point messaging plus full context reset and structured work logs externalize memory into auditable documents, directly targeting context contamination and long-running collaboration.
- Humans get a god's-eye view — read all messages, flip read state, impersonate any member — while each agent only sees its own unread inbox.
- Built-in local BPE token estimation and usage charts keep cost observable even when a model endpoint returns no usage field.
- A documented trial (7 agents, 1 hour, 126 messages, web Snake) shows the whole collaboration trace end to end.
- Hard dependency on the OpenClaw runtime (or a qclaw-style variant) that must be initialized first, adding an external layer and version-matching risk.
- Fully tested only on Windows 11 24H2; Linux/macOS are code-compatible but explicitly not guaranteed to work out of the box.
- The project's own fine-grained review style can turn small items into long discussions unless you loosen team.json and process rules.
- team RULE.md's T1–T5 actions and the scripts/ directory are marked as do-not-edit, limiting how far collaboration can be customized.
- Stable sessions need generous context budgets — the docs recommend contextWindow >= 128000 and maxTokens >= 8192 — which affects model choice and cost.
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 |
|---|---|---|---|---|---|---|
| CoordClaw Multi-Agent Collaboration System This agent | 42 · Major gaps | — | ★ 160 | 2mo ago | TypeScript | — |
| AgentChat | 29 · Major gaps | Self-hosted serviceFree + model costs | ★ 929 | 19d ago | Python | — |
| Event Planning Multi-Agent System | 62 · Some gaps | CLIFree + model costs | ★ 115 | 6mo ago | Python | — |
| AgentRQ | 64 · Some gaps | Self-hosted serviceFree + model costs | ★ 1.2k | today | Go | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
README describes a permission layering (human god's-eye view, agents only read their own unread messages, point-to-point messaging), an initial least-privilege design (least_privilege=1), but humans can impersonate any member identity and toggle any message read state without confirmation or audit constraints (user_confirmation=1). Data-flow transparency is decent: message flow, token consumption, and structured work logs are observable (data_flow_transparency=2). No dedicated sensitive-data handling; only local BPE token estimation is mentioned, no key/credential management policy (sensitive_data_handling=1). No dependency-security evidence: depends on OpenClaw/qclaw runtime, no dependency manifest or vulnerability mitigation (dependency_security=0). External effects: the setup wizard scans first-level subdirectories under the user home and may create junctions/symlinks, an operation with external side effects, though docs advise user judgment, so only 1 (external_effects=1). Rollback: only auto-reset and context reset mentioned, no rollback for executed external actions (rollback=1). Source attribution: README credits HUST AI Lab, but publisher is unverified by FollowAgents, so attribution is a claim only (source_attribution=1).
Self-consistency is good: README's T1–T5 actions, message loop, and context reset are corroborated by the bucketing/gating logic in the test file (self_consistency=2). Dependency availability is weak: hard dependency on OpenClaw or variants requiring prior environment initialization, Linux/macOS explicitly untested, dependency acquisition path unclear (dependency_availability=1). Failure messages: only test assertions and an LLM error-blocking switch, no user-facing error messaging design (failure_messages=1).
Audience and scenarios are well described: individual developers, team building, humanities/social-science research (audience_and_scenarios=2). Capability boundaries partially explained (single-agent insufficiency, collaboration threshold) but no systematic list of unsupported capabilities (capability_boundaries=1). Trigger precision: message loop and T1–T5 triggers described, but precise trigger conditions and boundaries lacking (trigger_precision=1). Environment fit: only Windows clearly supported, Linux/macOS untested, non-standard paths require manual findplatforms.json (environment_fit=1).
Information architecture is clear: README has complete sections covering install, usage flow, config tuning, tutorial links (information_architecture=2). Install notes are detailed with startup order and wizard (install_notes=2). Naming stability: teamsoul.md, team RULE.md, team.json are documented, but mixed Chinese/English and space-containing filenames reduce stability (naming_stability=1). Examples and tutorials are rich with illustrated guides and screenshots (examples_and_faq=2). Known limitations only scattered (Linux untested, non-standard paths), not consolidated (known_limitations=1). License is full MIT text (license=3). No version number or changelog (versioning_changelog=0). Maintenance responsibility: only Contributors and lab credited, no explicit maintainer or contact (maintenance_responsibility=1).
Output usability: structured work logs, message flow, and token charts make outputs reviewable (output_usability=2). Marginal value: multi-agent collaboration and context reset are differentiated, but evidence is a single Snake game example and value claims are largely asserted (marginal_value=1). Cost-benefit: emphasizes non-frontier models and local BPE estimation, but lacks cost comparison data (cost_benefit=1).
Claim traceability is weak: README makes many conceptual claims (probability is intelligence, difference is fuel) without verifiable sources (claim_traceability=1). Cross-source corroboration is limited: only the test file partially corresponds to README logic, no independent third-party corroboration (cross_source_corroboration=1). Fact/inference separation is insufficient: example results (1 hour, 126 messages) are self-reported without distinguishing measurement from inference (fact_inference_separation=1).
- Not found in source: dependency securityPin versions and run a dependency audit (npm audit, pip-audit) before installing; prefer running it in a container.
- Humans can impersonate any member identity and toggle message read states without confirmation or audit; use with caution in production.
- The setup wizard scans the user home directory and may create junctions/symlinks; verify target paths before running.
- Hard dependency on OpenClaw/qclaw runtime requiring prior initialization; Linux/macOS untested, cross-platform availability uncertain.
- No dependency manifest or vulnerability mitigation; dependency security cannot be statically confirmed.
- README claims are largely conceptual assertions and example results are self-reported without independent verification.