OpenSwarm is an open-source multi-agent AI system that acts as your entire office staff. Eight specialized agents collaborate through an intelligent orchestrator — research, write, analyze data, build slides, generate images, produce videos, and manage communications. All from a single prompt.
- Features
- Quick Start
- Usage
- Example Prompts
- Architecture
- Agent Reference
- Configuration
- OpenCode Integration
- Project Structure
- Development
- Custom Swarms
- License
OpenSwarm replaces an entire office toolchain with a team of AI agents that hand off work to each other automatically.
| Agent | What it does |
|---|---|
| Orchestrator | Routes every request to the right specialist. Breaks complex tasks into parallel subtasks and assembles final output. |
| Virtual Assistant | Email, calendar, Slack, file management, task tracking. 10,000+ integrations via Composio (Gmail, Slack, GitHub, HubSpot, Google Calendar). |
| Deep Research | Comprehensive web research with citations. Competitive analysis, academic paper search, evidence-based synthesis. |
| Data Analyst | Spreadsheet analysis, charts, statistical models — all inside an isolated IPython kernel. pandas, numpy, scipy, scikit-learn, matplotlib, seaborn, plotly. |
| Docs Agent | Formatted Word documents and PDFs. Full document lifecycle: create, view, modify, convert, restore. |
| Slides Agent | Polished HTML slide decks with speaker notes, exported to PPTX. Themes, layouts, images, charts, custom styling. |
| Image Agent | Generate and edit images via Gemini and fal.ai. Combine, edit, background removal. |
| Video Agent | Produce videos via Sora, Veo, Seedance. Edit audio, add subtitles, combine clips, trim. |
- Parallel execution — agents work simultaneously on split tasks
- File sharing — agents pass documents, images, slides between each other
- 10,000+ integrations — Gmail, Slack, GitHub, Google Calendar, HubSpot and more
- Multi-provider — OpenAI, Anthropic Claude, or Google Gemini
- Graceful degradation — missing API keys disable only the affected tools
- Auto-bootstrap — dependencies install themselves on first run
npm install -g @vrsen/openswarm && openswarmThe setup wizard handles everything: API keys, dependencies, configuration.
git clone https://github.com/musprodev/openswarm.git
cd openswarm
python swarm.pycp .env.example .env
docker-compose up --buildopenswarmLaunches the interactive terminal. Type a request and the orchestrator dispatches it to the right agents.
python server.pyFastAPI server on localhost:8080 with full API endpoints.
from swarm import create_agency
agency = create_agency()
agency.demo()Paste these into the terminal:
"Create a quarterly report from last quarter's data — analyze the numbers, build charts, write the report, make a slide deck"
"Research our top 5 competitors and write 3 SEO-optimized blog posts with images"
"Build a Q2 marketing campaign — strategy doc, creative assets, social posts, and a promo video"
"Schedule a team standup for tomorrow, send the agenda via email, and prepare meeting notes"
"Create an investor pitch deck with market research, financial projections, and executive summary" ┌─────────────┐
│ User │
└──────┬──────┘
│
┌──────▼──────┐
│ Orchestrator│
└──────┬──────┘
│
┌─────────────────┼─────────────────┐
│ Parallel │ Execution │
▼ ▼ ▼ ▼ ▼
┌──────────┐ ┌────────┐ ┌──────┐ ┌──────┐ ┌──────────┐
│ Virtual │ │ Deep │ │ Data │ │ Docs │ │ Slides │
│Assistant │ │Research│ │Analyst│ │Agent │ │ Agent │
└──────────┘ └────────┘ └──────┘ └──────┘ └──────────┘
┌──────────┐ ┌────────┐
│ Image │ │ Video │
│ Agent │ │ Agent │
└──────────┘ └────────┘
Each agent is an independent module with its own tools, instructions, and model config. The orchestrator handles task splitting, distribution, and result assembly. Agents communicate through a shared tool system and can pass files to each other.
| Agent | Entry point | Tools |
|---|---|---|
| Orchestrator | orchestrator/ |
Handoff, SendMessage |
| Virtual Assistant | virtual_assistant/ |
Email, Calendar, Slack, Files, Search |
| Deep Research | deep_research/ |
Web search, Scholar search |
| Data Analyst | data_analyst_agent/ |
IPython kernel, pandas, matplotlib |
| Docs Agent | docs_agent/ |
Create/Modify/Convert/View documents |
| Slides Agent | slides_agent/ |
HTML slides, PPTX export |
| Image Agent | image_generation_agent/ |
Generate, Edit, Combine, Remove BG |
| Video Agent | video_generation_agent/ |
Generate, Edit, Combine, Subtitles |
| Variable | Provider |
|---|---|
OPENAI_API_KEY |
OpenAI — GPT models + Sora video |
ANTHROPIC_API_KEY |
Anthropic — Claude models |
GOOGLE_API_KEY |
Google — Gemini models + Veo video |
| Variable | What it unlocks |
|---|---|
COMPOSIO_API_KEY |
10,000+ integrations (Gmail, Slack, GitHub, HubSpot) |
COMPOSIO_USER_ID |
User ID for Composio |
SEARCH_API_KEY |
Web search for Research Agent |
FAL_KEY |
Video editing, background removal |
PEXELS_API_KEY |
Stock photo search |
PIXABAY_API_KEY |
Stock photo search |
UNSPLASH_ACCESS_KEY |
Stock photo search |
DEFAULT_MODEL |
Override default model for all agents |
cp .env.example .envOr use the interactive wizard:
python onboard.pyOpenSwarm can be used as an agent inside OpenCode:
opencode --agent openswarmOr via opencode.json:
{
"agents": {
"openswarm": {
"description": "Multi-agent OFFICE AI team",
"path": "/path/to/openswarm/agents/openswarm.md"
}
}
}| File | Role |
|---|---|
agents/openswarm.md |
Orchestrator — entry point |
agents/openswarm-coder.md |
Software engineering |
agents/openswarm-data-analyst.md |
Data analysis & visualization |
agents/openswarm-deep-research.md |
Web research & synthesis |
agents/openswarm-docs-agent.md |
Document creation |
agents/openswarm-formatter.md |
Formatting & export |
agents/openswarm-image-agent.md |
Image generation |
agents/openswarm-plagiarism.md |
Originality checking |
agents/openswarm-reviewer.md |
Review & QA |
agents/openswarm-slides-agent.md |
Slide decks |
agents/openswarm-video-agent.md |
Video production |
agents/openswarm-virtual-assistant.md |
Email, calendar, Slack |
├── swarm.py # Agency config, agent imports, comm flows
├── server.py # FastAPI server
├── run_utils.py # CLI bootstrap & auto-install
├── onboard.py # Setup wizard
├── config.py # Model configuration
├── helpers.py # Integration helpers
├── shared_instructions.md # Cross-agent context
│
├── orchestrator/ # Orchestrator agent
├── virtual_assistant/ # Email, calendar, Slack, files
├── deep_research/ # Web research
├── data_analyst_agent/ # Data analysis & visualization
├── docs_agent/ # Document creation
├── slides_agent/ # Slide decks & PPTX export
├── image_generation_agent/ # Image generation & editing
├── video_generation_agent/ # Video generation & editing
│
├── shared_tools/ # Cross-agent tool system
├── patches/ # Runtime dependency patches
├── schemas/ # JSON data models
├── tools/ # State management, context
├── agents/ # OpenCode agent definitions
├── docs/ # Documentation
├── assets/ # Branding assets
│
├── bin/openswarm # npm entry point
├── docker-compose.yml # Docker deployment
├── Dockerfile # Docker build
├── .env.example # Environment template
├── .github/ # CI/CD & issues
└── AGENTS.md # AI agent customization guide
- Python ≥ 3.12
- Node.js ≥ 18
Optional:
- LibreOffice — enables slide thumbnails and previews
- Poppler / pdftoppm — enables PDF-based slide processing
git clone https://github.com/musprodev/openswarm.git
cd openswarm
cp .env.example .env
pip install -e .
npm installFork this repo and reshape it for any domain. Tell your coding agent:
"Turn this into an SEO optimization swarm."
The agent reads AGENTS.md and rewires everything automatically.
Ready-made ideas:
- SEO Swarm — keyword research, competitor analysis, blog writing
- Sales Swarm — lead research, outreach sequencing, proposal generation
- Marketing Swarm — campaign planning, creative assets, analytics
- Academic Swarm — research writing, formatting, bibliography
MIT — see LICENSE.