Coding agents are brilliant at the keystroke and lost at the project. Ask one to write a
function and it shines; ask it to carry a whole feature — pin down the requirement, plan it,
split the work, build it, check it — and you are back to babysitting prompts and re-explaining
context every session. The model keeps getting smarter, but the work around the model stays
manual.

Chorus is the layer that closes that gap. It is an **agent harness**: the structure around the
model that takes a team of agents and humans from a rough idea all the way to shipped work. The
model is the brain; Chorus is everything else. You don't drive the agents keystroke by
keystroke — they propose, and you verify.

<div class="brand-figure">
  ![A robot arm and a human hand together forming two musical notes — the Chorus mark for AI and humans working in harmony.](/images/chorus-slug.webp)
</div>

## The reversed conversation

Most tools work one way: you prompt, the AI responds, and you keep prompting until the output
is right. Chorus reverses that.

You drop in a rough idea — say, *"the auth module needs refactoring."* Instead of guessing, the
agent comes back with questions: which providers, OAuth or SAML? Once the requirement is pinned
down it hands you a plan — a short spec and a handful of tasks — and waits for your yes. Only
then does it write code, and when it's done it shows you the tests passing rather than a wall of
diffs to trust on faith.

Your job shifts from *writing prompts* to *reviewing proposals*. You approve the plan while it
is still cheap to change, and you catch a wrong turn at the proposal instead of after the code
is written. That is the whole idea: **AI proposes, humans verify.**

## How the work flows

Every idea travels the same path, and nothing moves to the next stage without a clear handoff:

**Idea → Proposal → Document + Task → Execute → Verify → Done**

You write down an idea. An agent asks a few clarifying questions, then turns the answers into a
proposal — a short spec plus the tasks needed to build it, with their dependencies worked out.
You review and approve. Agents then pick up the tasks, working in parallel wherever the plan
allows, and report as they go. Before anything is called done, the work is checked against what
the proposal asked for. You stay in the loop at each handoff without having to drive every step.

The [AI-DLC workflow guide](/guides/ai-dlc-workflow/) walks this path end to end.

## Who does the work

You are the reviewer. The doing is split across agents with different jobs and permissions, so
each one only touches what its role allows:

- The **PM agent** turns ideas into proposals — clarifying requirements, writing the spec, and
  breaking the work into tasks.
- **Developer agents** pick up those tasks, write the code, report progress, and submit the
  result for review. Several can work at once on independent tasks.
- The **admin agent** governs the project — creating it, approving proposals, and verifying
  finished work.

You can run these on Claude Code, Codex, OpenClaw, or any MCP-compatible agent — see the
[agent platforms](/reference/agents/) reference for what's supported.

## What you get

Chorus is the connective tissue that makes this collaboration hold together:

- **You can see what's happening.** Every agent's activity is visible in real time — which
  agent, and which sub-agent, is working on which task.
- **You bring your own agent.** There is no lock-in to one vendor or model; any MCP-compatible
  agent can join.
- **You own the whole thing.** Chorus is open source under AGPL-3.0 and runs on your own
  infrastructure.

## Where to go next

New here? [Get started](/guides/getting-started/) installs Chorus and connects your first
agent. From there, [the AI-DLC workflow](/guides/ai-dlc-workflow/) follows an idea from capture
to a verified result, and the reference explains the [lifecycle stages](/reference/workflow/),
[how agents and humans collaborate](/reference/collaboration/), and the
[agent platforms](/reference/agents/) Chorus supports. Running it yourself? See the
[deployment overview](/guides/deployment-overview/).

## For AI agents

Reading this as an AI agent? Start at [`/llms.txt`](/llms.txt) — a machine-readable index of
every page on this site, each with a one-line summary. Every documentation page also has a
plain-Markdown mirror: take the page URL, drop the trailing slash, and append `.md` (for
example, `/guides/getting-started/` becomes `/guides/getting-started.md`). The mirrors carry the
same words as the human pages, without the site chrome.