codeswarm
OfficialClick on "Install Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@codeswarmStart a swarm on issue #421 and open a reviewed PR."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
CodeSwarm Turn GitHub issues into tested, reviewed pull requests — autonomously. Self-hosted, multi-agent AI engineering swarm for GitHub. Provider-agnostic. Parallel. Production-ready.
Give CodeSwarm a GitHub issue. It turns it into a tested and reviewed pull request. Not AI that writes code — AI engineering team that ships production-ready PRs.
Table of Contents
Related MCP server: genswarms-fleet-mcp
Why CodeSwarm
Before | With CodeSwarm |
1 issue → hours of manual work | 1 issue → 5 AI agents → 1 ready-to-merge PR |
AI pair-programming | AI engineering team — GitHub-native, autonomous, self-hosted |
Wow effect: 1 GitHub Issue → 5 AI agents → parallel development → tests → review → CI fixes → 1 ready-to-merge PR
How it Works
flowchart TD
A[GitHub Issue<br/>ai:ready] --> B[Planner<br/>decomposes to JSON]
B --> C1[Researcher]
B --> C2[Coder]
B --> C3[Tester]
C1 --> D[Worktrees<br/>.codeswarm/task-xxx]
C2 --> D
C3 --> D
D --> E[Tester<br/>npm test / typecheck / lint]
E -->|fail| C2
E -->|pass| F[Reviewer<br/>security + correctness]
F -->|changes| C2
F -->|approved| G[GitHub PR]
G --> H{CI}
H -->|pass| I[Ready to merge]
H -->|fail| J[CI Agent<br/>parse logs → fix → push]
J --> HGitHub Issue (ai:ready)
↓ Planner → { goal, tasks: [...] }
├─ Researcher → finds files, git history, plan
├─ Coder ×N → isolated worktrees → commits
└─ Tester → npm test / lint / typecheck
↓ fail → Coder
↓ pass → Reviewer
↓ changes → Coder
↓ approved → GitHub PR → CI → CI Agent fixes on fail → ReadyFeatures
Parallel worktrees | Every agent works in isolated |
Conflict Resolver | LLM-assisted 3-way merge. Detects |
Autonomous PR loop | CI failure → parse logs → generate fix → push → re-run. Zero manual intervention. |
Provider-agnostic |
|
Cost & safety guards |
|
Full audit trail | Every |
Notifications | Telegram + Dashboard + GitHub comments. |
MCP native | Control swarm from Claude / Cursor / OpenCode via |
Quick Start
1. Install
npx codeswarm initCreates codeswarm.yml + .env. Fill .env:
DATABASE_URL=file:./data/codeswarm.db
GITHUB_TOKEN=ghp_xxx
GITHUB_OWNER=your-org
GITHUB_REPO=your-repo
OPENAI_API_KEY=sk-xxx
# or ANTHROPIC_API_KEY=sk-ant-xxx
# TELEGRAM_BOT_TOKEN=xxx
# TELEGRAM_CHAT_ID=xxx2. Run
Docker (recommended)
docker-compose up --build
# API: http://localhost:3000
# Dashboard: http://localhost:3001Local
npm install
npm run build
npm run dev
# or
node apps/api/dist/index.js &
node apps/dashboard/dist/index.js3. Trigger
Via label — add ai:ready to any open issue (webhook auto-creates task).
Via CLI — mock (no keys needed)
codeswarm issue run 421 --mock
# ✅ Local pipeline done: completed PR: https://github.com/example/repo/pull/438Via CLI — real
codeswarm issue run 421
codeswarm issue plan 421 # plan only, no codecodeswarm issue run first tries API (http://localhost:3000); if API is down, it falls back to local mock pipeline — great for demos and CI.
CLI Reference
codeswarm init # init codeswarm.yml + .env
codeswarm run --mock # start API + worker hint
codeswarm issue run <num> [--mock] [--repo owner/repo]
codeswarm issue plan <num> # planner only
codeswarm pr review <num>
codeswarm pr fix <num> # CI fix loop
codeswarm status # ASCII dashboard in terminal
codeswarm agents
codeswarm queue
codeswarm mcp # MCP server over stdioArchitecture
GitHub
│ webhook / poller
▼
Orchestrator ──► Task Queue ──► Agents (×N) ──► Git Worktrees ──► Tests / CI ──► GitHub PR
│ │ │
│ │ ├─ Planner / Researcher / Coder / Tester / Reviewer / CI Agent
│ │ └─ WorktreeManager + ConflictResolver
│ └─ InMemory (MVP) / Redis + BullMQ (prod)
└─ projects / tasks / agents / worktrees / artifacts / pull_requests / reviews / costs / agent_events
SQLite (MVP) → PostgreSQL (v0.3)Repo layout
codeswarm/
├── apps/
│ ├── api/ # HTTP API, webhook, orchestrator
│ ├── worker/ # queue worker (scale horizontally)
│ └── dashboard/ # real-time Web UI
├── packages/
│ ├── core/ # types, state-machine, cost, DB schema
│ ├── llm/ # OpenAI/Anthropic/Google/Ollama/LiteLLM/Mock
│ ├── agents/ # 6 agents + orchestrator
│ ├── github/ # Issues/PRs/Comments/Checks
│ ├── git/ # worktree manager
│ ├── queue/ # InMemory / Redis
│ ├── sandbox/ # security
│ ├── mcp/ # MCP server
│ └── telegram/ # notifier
├── tests/ # vitest (15 tests)
├── examples/ # wallet race condition
└── docker/MCP
Control the swarm from any MCP client:
{
"mcpServers": {
"codeswarm": { "command": "npx", "args": ["codeswarm", "mcp"] }
}
}Tool | Description |
| List tasks with status & progress |
| Start swarm for an issue |
| Cancel running task |
| Retry failed task |
| Spawn specific agent |
|
|
| Trigger reviewer |
Dashboard
http://localhost:3001 — live polling of http://localhost:3000/api/status
┌─────────────────────────────────────────┐
│ CODESWARM │
├─────────────────────────────────────────┤
│ ACTIVE AGENTS: 7 │
│ QUEUE: 12 │
│ PRs TODAY: 8 │
│ SUCCESS RATE: 81% │
├─────────────────────────────────────────┤
│ #421 Wallet race condition ██████ 80% │
│ #422 API timeout ████░░ 55% │
└─────────────────────────────────────────┘Terminal dashboard: codeswarm status
Configuration
codeswarm.yml (see codeswarm.yml.example):
task:
max_agents: 4
max_retries: 3
max_cost_usd: 2.00
timeout_minutes: 30
swarm:
max_total_agents: 10
max_depth: 2
llm:
provider: openai # openai | anthropic | openrouter | google | ollama | litellm
model: gpt-4o-mini
temperature: 0.2
queue:
type: memory # or redis
redisUrl: redis://localhost:6379
sandbox:
enabled: true
network: restricted
filesystem: repository_onlyCost example: gpt-4o-mini ≈ $0.15/$0.6 per 1M tokens. A full pipeline ~$0.01–0.05 with mock, ~$0.2–0.6 with real models.
Comparison
Tool | Positioning |
OpenHands | AI developer |
SWE-agent | Issue → code fix |
Aider | AI pair programming |
CodeSwarm | Orchestration of multiple AI developers — GitHub-native, multi-agent, parallel, autonomous, self-hosted, provider-agnostic |
Roadmap
Full roadmap with timeline, Gantt, metrics and exit criteria → ROADMAP.md
gantt
title CodeSwarm — Timeline
dateFormat YYYY-MM-DD
axisFormat %b %Y
section MVP
v0.1 shipped :done, v01, 2026-09-01, 2026-09-01
section Scale
v0.2 Parallel :done, v02, 2026-09-01, 2026-09-02
section Depth
v0.3 Plugins :done, v03, 2026-09-02, 2026-09-03
section GA
v1.0 Platform :done, v10, 2026-09-03, 2026-09-04Version | Status | Target | Highlights |
v0.1 | ✅ Shipped | Sep 2026 | Issues → Planner/Coder/Tester/Reviewer → PR, worktrees, OpenAI/Anthropic, Docker, SQLite, CLI |
v0.2 | ✅ Shipped | Sep 2026 | Parallel coders ×5, Redis/BullMQ, CI fix loop, Telegram, MCP, Dashboard SSE, cost guards |
v0.3 | ✅ Shipped | Sep 2026 | Conflict auto-merge, Postgres, multi-repo, vector memory, custom agents, plugin system |
v1.0 | ✅ Shipped | Sep 2026 | GitHub App, org RBAC, self-hosted models, marketplace, gVisor sandbox, analytics |
3 coders in parallel without file stomp
80% red CI auto-fixed, cost never exceeds
$2.00, p95 < 8 min visible in Dashboard
90% conflicts auto-merged, SQLite → Postgres via flag, 2 community agents contributed
Want to shape the roadmap? React with 👍 on issues labeledroadmap or open a roadmap/proposal.
Contributing
See CONTRIBUTING.md. PRs welcome!
npm install
npm run build # tsc -b
npm test # vitest — 15 testsLicense
MIT © 2026 CodeSwarm
This server cannot be installed
Maintenance
Resources
Unclaimed servers have limited discoverability.
Looking for Admin?
If you are the server author, to access and configure the admin panel.
Related MCP Connectors
Agent-native collaboration network: orchestrate a team of long-running agents from any MCP client.
MCP-first control plane for ProAgentStore agents and private instances.
Control plane for autonomous software labor. Agents claim objectives over MCP with audit trail.
Nifty's MCP server — exposes tasks, projects, messages, and files as tools for AI agents.
Related MCP Servers
- AlicenseNot gradedqualityAmaintenanceEnables MCP-compatible clients to connect to Swarm, providing access to Space memory, work coordination, artifacts, evaluations, context packs, and agent execution tools.14Apache 2.0
- FlicenseNot gradedqualityBmaintenanceMCP server that connects any MCP harness (Claude Code, claude.ai, Cursor) to a fleet of GenSwarms swarms for observation, configuration, and optional operation.
- AlicenseNot gradedqualityBmaintenanceEnables coordinating specialist agents through an event-driven backend, allowing submission of goals, retrieval of job status and results, and listing of jobs via MCP tools.MIT
- AlicenseNot gradedqualityAmaintenanceMCP server wrapping the taskswarm CLI as a single generic run tool for multi-agent task coordination.11MIT
Latest Blog Posts
- Who's Calling? MCP Hosts Are an Identity Blind Spot (And the Spec Knows It)By Om-Shree-0709 on .mcpAgent IdentityOAuth 2.1
- Your AI Chatbot Just Exposed Your CEO's Salary to an InternBy Om-Shree-0709 on .Agent IdentityMCP SecurityOAuth Delegation
- Why MCP Servers Need Execution Sandboxing (And Why Your Current Stack Isn't Enough)By Om-Shree-0709 on .Agentic AiPrompt InjectionWebAssembly
MCP directory API
We provide all the information about MCP servers via our MCP API.
curl -X GET 'https://glama.ai/api/mcp/v1/servers/aaxnet/CodeSwarm'
If you have feedback or need assistance with the MCP directory API, please join our Discord server