Open-source agent orchestration

Put a team of agents on the job.

The coordinator writes a plan and assigns each task to the agent that fits it. You approve the plan before any agent starts, then follow the run tree as the team works. When the work is done, it comes back to you for review.

See it run Run your first team View on GitHub

Run locally

git clone https://github.com/sabbour/agentweaver.git
cd agentweaver
npm run setup
npm run dev

Deploy to Azure

git clone https://github.com/sabbour/agentweaver.git
cd agentweaver
npm run azure:provision-infra

See the getting started guide for prerequisites and full details.

Agree on the plan before anyone starts.

The coordinator writes an OutcomeSpec that states what success looks like, what's in scope, what it assumes, and how you'll review the result. Agents start after you confirm the plan.

Confirmed Agentweaver Outcome plan beside the generated run tree — The plan you confirmed, beside the run tree it produced.

1
Define the outcome
Make success, scope, and constraints explicit.
2
Confirm the plan
Your confirmation starts the run.
3
Follow the graph
Watch dependencies, owners, models, and live status as the run moves.
4
Review the result
Review the team's work as one assembled result.

Cast a team for the job.

Start from a blueprint or build your own team. Each agent gets a role, a model, a charter, and the project skills it needs to do its part. The coordinator hands out the tasks, and you can always see which agent owns each one.

Explore team casting

Agentweaver specialist roster with the selected architect's model and skills — Each agent shows its role, model, and skills.

Agentweaver team casting proposal ready for review — Check the proposed team before you save it.

Drive it from any AI assistant.

Agentweaver ships an MCP server, so any MCP-capable client — GitHub Copilot CLI, Claude, Cursor — can create projects, cast teams, and start coordinator runs on your behalf. Prefer to stay in-app? The built-in Assistant calls the exact same tools, over the same MCP surface.

GitHub Copilot CLI

$ copilot mcp add agentweaver --transport http \
    --url https://<your-host>/mcp --header "Authorization: Bearer <token>"
✓ agentweaver added — 96 tools discovered
> Create a "Task Tracker" project and start the coordinator on it
● project_create({ name: "Task Tracker" })
  → project_id: 4b1a9e…  state: active
● coordinator_start({ project_id: "4b1a9e…", goal: "…" })
  → run_id: 9c2f31…  status: drafting
Drafted an OutcomeSpec for your review — nothing runs until you confirm it.

Auto-discovered from the repo's .mcp.json, or registered against a hosted deployment — see the MCP reference.

Assistant · Sessions

You What's blocking the Task Tracker run?

Assistant● run_status({ run_id: "9c2f31…" }) → status: awaiting_confirmation The plan's ready — it proposes Azure Container Apps for hosting. Want me to walk through the OutcomeSpec, or should I ask you to confirm it now?

No setup — every project has an Assistant session. See the Assistant runtime deep dive.

Every run keeps its work in one place.

Open a project and its board, runs, and results are already linked. Skills, memory, artifacts, telemetry, and cost all attach to the same run.

Northstar agent task board — Intake, active runs, review, and recovery on one board.

Project skills catalog with agent assignments — Assign reusable skills to the agents that use them.

Team memory with accepted decisions and pending proposals — Accepted decisions and open proposals stay with the project.

Trace every answer back to its model calls.

Reopen any run to see the evidence the coordinator used.

Live coordinator and child-agent streams
Per-agent model and AI-credit breakdowns
Persistent traces, files, decisions, and memory

See observability and cost

Agentweaver observability view with model mix and AI-credit usage

Ready to run your first team?

Get started Read how Agentweaver works