Run a team of AI agentsthat research, build, and ship for you.
Seventeen specialist agents, a mission board that turns goals into tasks, and the guardrails to trust what they do — open source, self-hosted.
From goal to finished work, hands-free
The orchestrator decomposes your goal into a task DAG, routes each step to the right agent, and pauses for your approval wherever you want a checkpoint.
Everything you need to run AI agents in production
A full production stack — streaming chat to Kubernetes deployment. Real tools, real guardrails, no toy demos.
Talk to every agent in one place
Talk to any agent in the fleet from one streaming chat. Tool calls show up live as they run — search queries, SQL, Slack posts — so you see what the agent actually did, not just its final answer.
- Streaming responses
- Live tool-call cards
- Switch agents mid-conversation
Organize work with a live task board
Give the fleet a goal and the orchestrator turns it into a Kanban DAG — a chain of steps agents pick up, execute, and hand off. Human-in-the-loop checkpoints pause the run wherever you want a say.
- Goal → step DAG
- Human-in-the-loop checkpoints
- Live mission board
Describe your goal
Type what you want in plain language — a report, a campaign, a migration. No workflow to configure, no DAG to draw by hand.
Auto-execute the DAG
The orchestrator breaks the goal into dependent steps and hands each one to the specialist agent best suited to run it.
Track & intervene
Watch every task move across the board live. Approve checkpoints, retry a failed step, or step in anywhere the run needs a human call.
See every step, token, and cent
Every step is traced in Langfuse — latency, token usage, and cost per tool call, live as the run happens. When a run is slow or expensive, you see exactly which step, not just an aggregate number.
- Per-step latency + cost
- Langfuse trace timeline
- p95 94ms under load
Create new agents without redeploying
Compose a new agent at runtime — pick a system prompt, a model, and wire in external MCP tools — without touching code or redeploying.
- No redeploy
- MCP tool wiring
- Model per agent
Ground answers in your own documents
Upload your own docs and every agent can search them through local pgvector retrieval — chunked and embedded on your machine, grounding answers in your content instead of the model's guesses.
- Local pgvector search
- fastembed, on-device
- Grounded citations
Ship agents as APIs and MCP servers
Ship an agent behind a versioned API key, embed it as a widget, or expose your whole fleet as an MCP server. Every publish is a version — roll back a bad one in one click.
- Versioned publishing
- One-click rollback
- MCP server export
Chain agents into automated pipelines
The orchestrator sequences the steps, hands each one to the right agent, and runs the whole chain unattended.
Dependency Chains
Steps declare what they depend on. The orchestrator won't start a step until everything it needs has finished.
One-Click Execution
Launch a whole pipeline from one goal — no manual hand-off between steps, no copy-pasting output between agents.
Visual Task Board
Every pipeline run shows up as tasks on the mission board, so you can watch, approve, or retry any step as it happens.
Specialized agents, ready to deploy
Each agent ships preconfigured with the right tools and prompts. Use the built-in lineup or create your own.
17 agents total — real system prompts, real tools, no stubs. Sixteen run on LangGraph; Fact Checker runs on Pydantic AI, to prove the platform isn't locked to one framework.
The governance layer that makes agents safe to ship
Most agent demos stop at a chat window. Evals, guardrails, cost caps, and versioned rollouts are what separate a production system from a toy.
Ready to put agents to work?
Self-host it in an afternoon. MIT licensed, docker-composed, ready to extend.