Self-hosted · Open source

Orchestrate your
AI coding agents

Autonomous Kanban runner for Claude Code, Codex, Cursor, Kimi and more. Coordinate parallel agent waves from a single dashboard.

Open Board →
Claude Code OpenAI Codex Cursor CLI Kimi Custom agents MCP
Features

Everything you need to run agent teams

Autonomous Runner

Launches each agent's CLI as a subprocess, pulls the next task from its queue, captures output, and advances the board — fully unattended.

🌊

Wave Orchestration

Group tasks into dependency-ordered waves. Agents run in parallel within each wave; the next wave starts only when dependencies clear.

🔍

Peer Review Gate

Assign a reviewer agent per task. The runner dispatches the reviewer automatically after the owner finishes — changes_requested loops back for a fix.

🧠

Semantic Memory

pgvector + Ollama embeddings store episodes, lessons and playbooks. Low-scoring tasks auto-generate lessons injected into future agent briefs.

🛠️

24 MCP Tools

Full MCP server: board, tasks, runner, git, deps, tests, shell, release, CSV export, time log, risk register, decisions, announce, digest.

📋

Actions Tab

Milestones, time tracking, risk register, ADR decisions, approval gates, Telegram/webhook announcements, and a full DevOps panel.

🌿

Worktree Isolation

Each task runs in its own git worktree+branch. Auto-merges into main on Done, serialised to avoid conflicts. Conflict cards are flagged automatically.

📁

File Upload & Chat

Upload PDFs, DOCX, XLSX — text extracted and stored in the DB. Ask agents questions about your project documents directly from the board.

🔒

Auth & Multi-user

Bearer token + Google OAuth. Allowlist-based access control managed live from the UI — no restart needed when adding users.

How it works

From task to merged commit

1

Create a project & import tasks

Define agents with their CLI templates, strengths and WIP limits. Import tasks from markdown, GitHub issues, or create them manually.

2

Auto-allocate and plan waves

Maestro scores each task against agent strengths, file-collision risk and token budget, then assigns and orders waves automatically.

3

Hit Run — agents execute in parallel

The autonomous runner starts each agent's CLI in its own worktree. Output is captured, tokens tracked, and the board updates in real time.

4

Review, merge, repeat

Reviewer agents gate Done cards. On approval, the worktree branch merges into main automatically. Lessons from failures feed the next wave.

Built on boring, reliable tech

FastAPI
Python backend, async, single file
PostgreSQL 16
19-table schema, psycopg3 native
pgvector
Semantic memory, 768-dim HNSW
Ollama
Local embeddings, nomic-embed-text
Caddy
TLS termination, reverse proxy
Vanilla JS
Single-file frontend, no build step
24
MCP tools
19
DB tables
5+
Agent vendors
0
Build steps