Agentweaver

Appearance

AgentweaverRun a team of AI agents for any scenario you can describe.

Submit a goal. A coordinator drafts a confirmed plan. Named specialist agents work in parallel, each in their own sandbox. You watch every step live, steer mid-run, and review before anything merges.

What is Agentweaver?
Get started
Deploy to AKS
MCP server
Agentweaver

Coordinator orchestration

Submit a goal. The coordinator drafts an OutcomeSpec, you confirm it, then a squad of specialists works in parallel — each in an isolated git worktree. You steer mid-run, review once, and approve.

Full observability

Every agent message, tool call, and tool result streams live over SSE and is persisted before fan-out. Nothing is opaque. Reopen any run and see the full audit trail.

Human gates that matter

No code merges without your explicit approval. The OutcomeSpec gate, RAI check, and human review step are enforced by the platform — not optional.

Scenario workflows

Seven built-in YAML workflows cover software delivery, bug fixing, code review, content authoring, PM discovery, incident response, and agent evaluation. Add your own or generate one.

Persistent team memory

Agents build on prior work through four memory layers — decisions, core context, learnings, and open session. A Scribe records what the team learned after every run.

Sandbox browser preview

When an agent starts a dev server inside its isolated Kubernetes sandbox pod, open a live browser preview of it from the run view — a port-forward tunnel scoped to that one run's pod, with no egress widening. See it in action →

AI credit and token usage monitoring

Track GitHub Copilot token consumption and AI Credits at every level — individual runs, workflow orchestrations, projects, and the entire app. Live counters during execution, dashboards for analysis.

See it in action

Works the way you do

Self-hosted. Runs with GitHub Copilot or Microsoft Foundry. Drive it from the web UI or any MCP-compatible client, including Copilot CLI.

A Copilot agent in every project

Each new project ships with a ready-to-use GitHub Copilot agent under .github/agents/ that drives Agentweaver through its MCP tools. Its tool map is generated from the live tool set, so it never drifts — and it never overwrites your edits. See it in action →

Alpha software

Agentweaver is alpha software under active development. Expect breaking changes, incomplete features, and rough edges. Do not rely on it for production workloads.

How it works ​

Submit a goal. The coordinator drafts a plan — you confirm it before any work starts. A squad of specialists works in parallel, each in an isolated sandbox. Review the assembled work once, behind a single gate that includes a Responsible AI check.