agent-swarm
About
agent-swarm is an open-source (MIT) orchestration layer that runs a persistent team of AI coding agents in isolated Docker containers with shared, compounding memory.
Details
- Author
- desplega-ai
- Categories
- Developer Tools
Jump to
Setup
Install agent-swarm in your MCP client (Claude Desktop, Cursor, Windsurf, and others).
Repository: https://github.com/desplega-ai/agent-swarm
Follow the installation instructions in the repository README, then restart your MCP client.
An engine to make your company AI Native
Built bydesplega.sh— by builders, for builders.
[!TIP]This repo evolves every single day.Watch now →
▸daily evolution·slack → pr·Making of
agent-swarm.dev is an open-source operating system for AI work: a lead agent breaks goals into tasks, routes them to specialized workers such as Claude Code or Codex, runs each worker in an isolated container, and preserves shared memory, tools, schedules, and review gates so delegated work compounds across sessions.
AI-Native · Compounds · Presence · Harness & LLM-Agnostic · Your Infra · Your Memory ·
agent-swarm.dev runs a team of AI agents that coordinate autonomously. Alead agentreceives tasks (from Slack, GitHub, GitLab, Linear, Jira, email, or the API), breaks them down, and delegates toworker agentsrunning in isolated environments (Docker). Workers execute tasks, ship solutions, and write their learnings back to a shared memory so the whole swarm gets smarter every session.
You can run agents for Marketing, Product, UX, Engineering, Support, Operations, HR, Finance, or any role you can think of. A centralized Lead coordinates them, and they share the learnings horizontally. That's the true difference betweenAI FirstandAI Native.
agent-swarm.dev is the shared cloud brain and muscle that makes your whole company better every day.
Sometimes humans are the blocker. We can help you. Contact uscontact@desplega.sh.
Learn more in thearchitecture overview.
flowchart LR subgraph IN["Tasks come in"] direction TB S["Slack"] G["GitHub / GitLab"] E["Email"] A["API / CLI"] end LEAD(["Lead Agent<br/>plans & delegates"]) subgraph WORKERS["Workers in Docker"] direction TB W1["Worker"] W2["Worker"] W3["Worker"] end subgraph BRAIN["Persistent brain"] direction TB MEM["Memory<br/>(vector search)"] ID["Identity<br/>(SOUL, CLAUDE.md)"] end subgraph OUT["Work ships"] direction TB PR["Pull Requests"] REPLY["Slack replies"] EMAIL["Email replies"] end IN --> LEAD --> WORKERS WORKERS -->|reads context| BRAIN WORKERS -->|writes learnings| BRAIN WORKERS --> OUT
Use cases that are used daily by ourselves and others. Each playbook contains: the agents, the tools & skills, and workflows & schedules behind it.Browse all playbooks →
- Feature Development— Integrated with Linear and GitHub to take feature requests from Slack and turn them into pull requests.
- Lead Prospecting— Integrate your prospecting tools with the swarm and let agents handle outreach, scheduling, and follow-up.
- Content Generation— Generate engagement tools, blog posts, manage social media presence, update your website, and more.
- UX Command Center— Agents that keep your product usable: record agentic sessions, enforce your design system, and mine user logs to detect and propose UX improvements.
- Proactive Customer Support— Agents that oversee your top accounts, prepare scheduled reports, and leverage everything they know about your platform to keep those accounts up to date.
- Code Health & Alert Management— Datadog, New Relic, Sentry, or any alerting tool can kick off fixes or new proposals. Monitor code health and propose improvements weekly, daily, or hourly.
- Reports from Multiple Sources— Integrate your data warehouse to generate tailored reports and answer the key questions your team has, with fresh data. Your BI tool may be a thing of the past.
- Self-Documenting & Release Reports— Update your docs and use frameworks likeRemotion,qa-use, andbrowser-useto generate release videos and rich documentation in seconds, at the cadence you need.
- Do you have a cool playbook to share? Send us a PR!
The patterns that compound.Five recipes show up in nearly every playbook — they're how the swarm stays reliable as it scales:Litmus Tests(LLM-as-judge quality gates) ·Drain Loops(one ticket → a chain of reviewable PRs) ·HITL Gates(pause for human approval on irreversible steps) ·Per-Customer Working Directories(context that compounds per account) ·No-op Workflows(skip silently when nothing changed).See all patterns →
- Lead/worker orchestration in Docker— isolated dev environments, priority queues, pause/resume across deploys, and API-side alerts when claimable work stops being picked up.Architecture →
- Compounding memory & persistent identity— agents remember past sessions and evolve their own persona, expertise, and notes, with ratcheting file budgets and visible recovery when local edits exceed prompt limits.Memory →·Agents →
- Hybrid + graph-linked memory recall— memory retrieval can blend vector and full-text ranking, expand through linked memories, surface usefulness readouts, and let agents correct an existing memory without losing its ID or history.Memory →·MCP tools →
- Multi-channel inputs— Slack, GitHub, GitLab, email, WhatsApp, Linear, Jira, and the HTTP API all create tasks.Integrations
- Persistent Slack thread trees— opt in to one editable task tree per conversation, complete streamed outcome cards, and explicit agent-authored Block Kit messages without routine relay spam.Slack guide →
- Workflow engine with Human-in-the-Loop— DAG-based automation with approval gates, retries, structured I/O, andforeachfan-out that rejoins child agent tasks deterministically.Workflows →
- Scheduled & recurring tasks— cron-based automation for standing work, with schedules that can target agent tasks, workflows, or catalog scripts.Scheduling →
- Mid-run task steering— add context at the next turn boundary, request an immediate interrupt where the harness supports it, or degrade safely to a follow-up task.Task steering →
- Operator configuration UI— tune non-secret feature flags, limits, integration toggles, and runtime defaults from Settings → Configuration, with source and restart-required indicators.Configuration →
- Asset namespaces— group tasks, workflows, schedules, pages, apps, scripts, and mapped files under canonical shared or personal keys, with inheritance, subtree filtering, audited moves, and cross-entity discovery.Asset namespaces →
- Persistent shared files via agent-fs— co-deploy agent-fs for shared task attachments, previews, and search; late provisioning or config reloads activate the provider without an API restart, and tenant-authenticated control planes can invite shared-org members without receiving the bootstrap key.Co-deployment guide →
- Durable script workflows— launch background script runs, inspect their journals, and track them from the dashboard when a one-shotscript-runis too small.Guide →
- Scripts-only MCP mode (code-mode)— reduce the external MCP surface to eight script tools while retaining the full swarm SDK throughscript-run, cutting tool-schema context for capable coordinator models.Guide →
- Scripts as external APIs— expose a saved script as a publicPOST /api/x/script/<id>endpoint with optional bearer auth, typed input validation, and per-endpoint usage tracking.Guide →
- Typed script API connections— lead-managed OpenAPI, GraphQL, and MCP connections generatectx.api./ctx.mcp.clients for scripts, with credential bindings and OAuth-backed auth kept server-side.Guide →
- Swarm Apps— agents build versioned, schema-backed dashboard apps with reusable UI elements, named queries and actions, per-user settings, RBAC, history, and safe rollback. Models can also sync source-backed rows through owner-scoped script connections.Apps API →
- E2B-backed eval harness— run a scenario × harness-config matrix against real swarm stacks, capture transcripts/artifacts, and grade outcomes with deterministic checks plus LLM or agentic judges.Guide →
- Harness & LLM agnostic— run with Claude Code, Claude Bridge, OpenAI Codex, pi-mono (Anthropic, OpenRouter, or Amazon Bedrock), Devin, Claude Managed Agents, raw LLMs, or opencode. The dashboard model picker follows a live models.dev catalog (with a bundled snapshot fallback), so newly released provider models appear without a redeploy. Route every OpenRouter-backed harness, workflow, and summarizer through an OpenAI-compatible gateway withOPENROUTER_BASE_URL. Tasks, schedules, and workflow agent-task nodes can use portablemodelTierintent (smol,regular,smart,ultra), and operators can set per-agent reasoning effort (off→max, where supported) without changing task payloads.Harness config →·Add a new provider →
- Published release artifacts— every release publishes multi-architecture API, full-worker, and slim-worker images alongside versioned E2B templates, the npm CLI package, and the Helm chart.Artifact inventory →
- OpenTelemetry traces plus OTLP cost/token metrics— export API + worker traces and finalized session cost/token counters through the same OTLP pipeline for dashboarding in SigNoz, Datadog, Tempo, or another compatible backend.Observability →
- Follow-up continuity across all harnesses— child tasks inherit a bounded prior-task context preamble built from the task chain, so continuity survives restarts and works the same across every provider.Task lifecycle →
- Skills & MCP servers— reusable procedural knowledge, bundled skill reference files, and per-agent MCP servers with scope cascade.MCP tools →
- External tool-router access— thexcommand andswarm_xMCP tool let humans and agents execute approved third-party routes such as Composio without baking bespoke MCP servers first.CLI →·Composio →
- Config-driven metrics dashboards— define read-only SQL widgets, version them, and render them in the dashboard without shipping custom frontend code.Metrics API →
- DB-backed pages— agents publish HTML or JSON pages (reports, dashboards, action specs) viacreate_page, remove stale pages withdelete-page, and share them with public / authed / password modes, version history, view counters, diff helpers, and PDF export.MCP tools → Pages
- KV store— Redis-like namespaced key/value store with auto-scoped context (Slack thread / PR / Linear issue / page).MCP tools → KV
- Real-time, themeable dashboard + task attachments— monitor agents, tasks, per-user usage costs, and inter-agent chat; choose from built-in light/dark themes; filter tasks by requester; create tasks with uploaded files; and preview attachments inline above session prompts.app.agent-swarm.dev →
Need help? Contact us atcontact@desplega.sh.
Prerequisites:Dockerand at least one supported harness credential. The default quick start assumes aClaude CodeOAuth token (claude setup-token), but pi-mono / Bedrock, Codex, Devin, and other provider setups are also supported.
The fastest way is the onboarding wizard — it collects credentials, picks presets, and generates a workingdocker-compose.yml:
bunx @desplega.ai/agent-swarm onboard npx @desplega.ai/agent-swarm onboard
Prefer manual setup? Clone and run with Docker Compose:
git clone https://github.com/desplega-ai/agent-swarm.git cd agent-swarm cp .env.docker.example .env # edit .env — set API_KEY plus the credential for your chosen harness (for example CLAUDE_CODE_OAUTH_TOKEN) docker compose -f docker-compose.example.yml --env-file .env up -d
The API runs on port3013, with interactive docs athttp://localhost:3013/docsand an OpenAPI 3.1 spec athttp://localhost:3013/openapi.json.
- Local API + Docker workers— run the API on your host, workers in Docker. SeeGetting Started.
- Claude Code as the lead agent—bunx @desplega.ai/agent-swarm connect(ornpx @desplega.ai/agent-swarm connect), then tell Claude Code to register as the lead.
You (Slack / GitHub / Email / CLI) | Lead Agent ←→ MCP API Server ←→ SQLite DB | ┌────┼────┐ Worker Worker Worker (Docker containers with full dev environments)
- A task arrives via Slack DM, GitHub @mention, email, or the API.
- The lead plans and delegates subtasks to workers.
- Workers execute in isolated Docker containers (git, Node.js, Python, etc.).
- Progress streams to the dashboard, Slack threads, or the API.
- Results ship back out as PRs, custom pages, issue replies, or Slack messages.
- Session learnings are extracted and become memory for future tasks.
More detail in thetask lifecycle docs.
Real-time monitoring of agents, tasks, and inter-agent chat. The task list can be filtered by requester, includingMeandUnattributed, while keeping the selection in the page URL. Session titles can be renamed inline, and agent profiles can search hundreds of Lucide icons, pick a color, and retain deterministic defaults when no customization is set. Use the hosted version atapp.agent-swarm.dev, or run locally:
cd apps/ui && bun install && bun run dev
bunx @desplega.ai/agent-swarm <command> npx @desplega.ai/agent-swarm <command>
For production deployments (Docker Compose with multiple workers, systemd for the API, graceful shutdown, integration config), seeDEPLOYMENT.mdand thedeployment guide.
Everything lives atdocs.agent-swarm.dev. Good starting points:
- Getting Started— install, configure, and run your first task
- Architecture overview— how the swarm is wired together
- Playbooks— eight production flows we use to run Desplega, plus the patterns behind them
- CLI referenceandEnvironment variables
- API reference— every HTTP endpoint
We love contributions! Whether it's bug reports, feature requests, docs improvements, or code — all are welcome.
Join ourDiscordif you have questions or want to discuss ideas.
Are you an agent? Go toagent-swarm.dev/skill.md.
The charts are generated from timestamped GitHub stargazer data and refreshed weekly through a reviewable pull request.
This is a web browser that enables your coding agent, such as Claude Code, to visit websites on your behalf and assist you in identifying bugs or creating UI test cases.
Create crafted UI components inspired by the best 21st.dev design engineers.
Bring agent evaluations, observability, and synthetic test set generation directly into your IDE for free with Galileo's new MCP server
An MCP server to help AI assistants to answer questions and generate AccelByte Extend SDK code more effectively .
MCP server for AI Diagram Maker — generate beautiful software engineering diagrams directly inside Cursor, Claude Desktop, Claude Code, or any MCP-compatible AI agent
ALAPI MCP Tools,Call hundreds of API interfaces via MCP
AI-powered SVG animation generator that transforms static files into animated SVG components using the Allyson platform
MCP server that gives AI assistants on-demand access to 1,500+ amCharts docs, ~300 code examples, and 1000+ class API references.
APIMatic MCP Server is used to validate OpenAPI specifications using APIMatic. The server processes OpenAPI files and returns validation summaries by leveraging APIMatic’s API.
One shared context layer for AI agents and humans — live API specs, DB schemas, and versioned contracts across repos so every agent and teammate works from the same source of truth.
Build and deploy full-stack Next.js apps with 98 tools for React, AWS, and MongoDB
Sign in to leave a review
Use Google, GitHub, or an email account so ratings stay tied to real people.
No reviews posted yet.
