SwarmClaw
A self-hosted control plane for running, orchestrating, and supervising agent teams with memory, tools, and schedules.
Per-dimension scores and reasoning
Evidence shows some permission controls (e.g., reset parameter in MCP server test route), but no explicit user confirmation mechanism or data flow transparency documentation. Sensitive data handling (e.g., credential encryption) is mentioned in changelogs but lacks detailed implementation. Dependency security has overrides and CI but no vulnerability scanning evidence. External effects (e.g., connectors) exist but are not explicitly limited. Rollback mechanisms (e.g., guardian recovery) appear in test files but are not documented. Source attribution (e.g., author info) exists but publisher is unverified.
Self-consistency is good: README, package.json, and CI config align, version numbers match. Dependency availability: package.json lists many dependencies but no lockfile or integrity checks. Failure messages: changelogs mention improved error messages but no concrete examples.
Audience and scenarios are clear (personal assistant, virtual company, dev team), but capability boundaries are not explicitly defined. Trigger precision: schedules and task systems are described but no precise trigger conditions. Environment fit: supports multiple deployment methods (desktop, npm, Docker) but not all platform-specific requirements.
Information architecture is clear (README has TOC), install notes are detailed (multiple methods), naming is stable (version numbers consistent), examples and FAQ are provided, known limitations are not explicitly listed, license is MIT, versioning changelog is detailed, maintenance responsibility is unclear (publisher unverified).
Output usability: provides multiple use cases and screenshots but no actual output examples. Marginal value: as an alternative to Claude Code and LangChain, has unique features (e.g., memory, orchestration) but no comparison with others. Cost-benefit: no pricing or resource consumption info.
Claim traceability: README feature claims are partially reflected in code and tests but not fully. Cross-source corroboration: CI and test files support some claims but lack independent verification. Fact-inference separation: documentation does not clearly distinguish facts from inferences.
- Publisher is unverified; treat source with caution.
- Many dependencies without lockfile; supply chain risk needs attention.
- Permission control details are unclear; review actual code.
- Rollback mechanism appears only in tests, not documented.
What does this agent do, and when should you use it?
SwarmClaw is a self-hosted AI agent runtime and multi-agent framework with a dashboard, task board, org chart, and chat interface. It runs heartbeats, schedules, background jobs, delegated tasks, and restart-safe orchestration while retaining conversations, memory, and run state. Built-in provider paths include Claude Code, Codex, OpenCode, Gemini, OpenAI, Anthropic, OpenRouter, Ollama, and compatible endpoints. It connects MCP servers, messaging connectors, and OpenClaw gateways, with desktop, global npm, Docker, and hosted-image deployment options. It fits teams that need persistent, multi-role automation under their own control plane rather than a one-off chat integration.
Running swarmclaw starts a server at http://localhost:3456, where operators manage agents, chats, tasks, schedules, and connectors. Agents can use built-in tools and stdio, SSE, or streamable HTTP MCP tools configured from the MCP Servers panel; they can delegate to Claude Code, Codex CLI, OpenCode, Gemini CLI, Cursor Agent CLI, Qwen Code CLI, or native SwarmClaw subagents. Orchestration supports branches, loops, parallel paths, explicit joins, and restart-safe state; outputs can persist as chat history, task outputs, structured-session transcripts, or connector deliveries. Scheduled runs record lastDeliveryStatus and lastDeliveryError, and a failed dead-lettered task can be requeued through POST /api/tasks/:id/retry.
- An operations lead builds an agent org chart and delegates ongoing work to specialized development, research, and operations roles.
- An engineering team has a builder agent invoke Codex CLI, Claude Code, or OpenCode while testing and review agents track work on the task board.
- A research group schedules daily or weekly scans, has several agents investigate independently, and consolidates their findings into a cited report.
- A support team needs agents to receive and send messages through Discord, Slack, Telegram, WhatsApp, Teams, or email.
- An OpenClaw operator manages several gateways, agent files, skills, and remote runtimes from one control plane.
- An on-chain operations team needs recurring balance checks, transaction simulation, or approval-threshold controls for agent-linked wallets.
What are this agent's strengths and limitations?
- Combines chats, tasks, org structure, schedules, memory, and restart-safe orchestration in one self-hosted control plane.
- Natively spans cloud APIs, CLI backends, local Ollama/LM Studio runtimes, OpenClaw, and compatible custom endpoints.
- Supports stdio, SSE, and streamable HTTP MCP servers, with per-agent configuration, testing, and assignment.
- Provides operational evidence for scheduled work through delivery status, credential preflight, retry backoff, and orphan-task recovery.
- Structured Sessions include templates, participants, durable transcripts, outputs, and a visual flow builder.
- Local and self-hosted operation requires Node.js 22.6+; Docker Desktop is recommended for sandbox browser execution.
- API providers need their own credentials, while delegated CLI backends require the relevant provider CLIs to be installed separately.
- Self-hosted deployments require the operator to manage persistent storage,
ACCESS_KEY,CREDENTIAL_SECRET, network access, and connector configuration. - The macOS desktop app may need a manual Gatekeeper workaround when signing or notarization is unavailable; the related issue remains open.
- Wallets, messaging connectors, and external tools broaden permission and operational risk, so approval limits and runtime guardrails need configuration.
How do you install or deploy this agent?
Node.js 22.6+ and npm 10+ are required. For a global install: npm i -g @swarmclawai/swarmclaw, then run swarmclaw. From source: git clone https://github.com/swarmclawai/swarmclaw.git, cd swarmclaw, nvm use, and npm run quickstart. For Docker, from the repository run mkdir -p data, touch .env.local, and docker compose up -d --build, then open http://localhost:3456. Hosted deployments need persistent storage mounted at /app/data plus ACCESS_KEY and CREDENTIAL_SECRET; API-based providers also need their corresponding credentials configured in provider settings.
How do you use this agent?
After startup, open http://localhost:3456, create an agent, and select a configured provider and model. Assign tools, memory, runtime skills, or MCP servers, then begin work from chat or the task board; configure delegation or Structured Sessions for collaboration. Create schedules for recurring jobs and inspect their saved delivery status and run records. For OpenClaw, configure a gateway profile and select its runtime from the control plane; to install the OpenClaw skill, run openclaw skills install swarmclaw.
How does this agent compare with similar options?
The project describes itself as a practical alternative to Claude Code and LangChain. Compared with a single coding CLI, it emphasizes multi-agent delegation, tasks, schedules, memory, and connectors; compared with a framework-oriented orchestration library, it supplies a deployable control plane, runtime, and operator UI. It can still use Claude Code and Codex CLI as delegated backends.
FAQ
Is it tied to one model provider?
Can I trace a failed scheduled task?
lastDeliveryStatus and lastDeliveryError; connector delivery uses an outbox with retries, backoff, and per-run deduplication.