Chorus

About

Chorus is an AGPL-3.0 open-source agent harness — the layer wrapped around an LLM agent that owns session lifecycle, task state, subagent orchestration, observability, and failure recovery. In its own words, the model is the brain and Chorus is everything else. It organizes work as an AI-DLC pipeline — Idea → Proposal → Document + Task DAG → Execute → Verify → Done — inspired by the AI-Driven Development Life Cycle methodology from AWS. Its core idea is the Reversed Conversation: instead of a human prompting and the AI executing, the AI proposes and the human sits in the verify seat. Multiple agents and multiple humans collaborate in one project, governed by a 5-resource × 3-action permission matrix (presets plus custom combinations) with no fixed roles. Agents connect over MCP (HTTP Streamable Transport) through a full set of permission-gated MCP tools; Claude Code, Codex CLI, Kiro CLI, Pi and OpenClaw are officially supported, an OpenCode plugin is community maintained, and any other MCP agent can use the standalone SKILL.md that ships with the project. chorus daemon turns a machine into a long-running agent runtime: one daemon serves many working directories via repeatable --cwd flags, every (agent, host, cwd) combination is an individually addressable instance, and it can be installed as a service that starts at boot. Self-hosting is a single line — npm i -g @chorus-aidlc/chorus && chorus — with embedded PGlite, so no database, no Docker and no configuration; it listens on http://localhost:8637 by default. For production you can switch to external PostgreSQL + Redis via Docker Compose, or deploy on AWS with the bundled CDK stack. The current release is v0.16.0, and the UI ships in English, Chinese, Korean and Japanese with light / dark / system themes.

Key Features

• ✓AI-DLC pipeline: Idea → Proposal → Document + Task DAG → Execute → Verify → Done
• ✓Reversed Conversation: the AI proposes and the human holds the verify seat
• ✓Requirement elaboration: structured Q&A rounds run before any proposal is written
• ✓Proposal approval flow: a PM drafts, an Admin approves, and drafts materialize into real documents and tasks
• ✓Task DAG: dependency modeling, cycle detection and interactive visualization
• ✓Two-pass acceptance criteria: dev agent self-check plus admin review, with pass / fail evidence per criterion
• ✓Real-time Kanban: tasks flow To Do → In Progress → To Verify automatically, with live agent presence
• ✓Project resource mind map: Ideas / Proposals / Documents / Tasks as one collapsible live tree
• ✓Fine-grained agent permissions: a 5-resource × 3-action matrix with presets and custom combinations
• ✓chorus daemon: a long-running local agent runtime where one daemon serves many working directories and each instance is addressable
• ✓Multi-client access: Claude Code, Codex, Kiro CLI, Pi, OpenClaw and any other MCP agent
• ✓Cmd+K global search: 6 entity types across 3 scope levels, sharing one backend between web UI and agents
• ✓One-line self-hosting: embedded PGlite needs no database or Docker, with Docker Compose or AWS CDK for production
• ✓Reference artifacts: attach docs, repos, issues and articles to any idea, proposal or task, readable from web and MCP
• ✓Session lifecycle and audit: heartbeats, auto-expiry, failure recovery and a full activity trail