Run Claude, Codex, Gemini, Antigravity, Kimi, OpenCode, and Grok against your code in parallel — through the CLI subscriptions you already pay for. Multiple reviewers, no extra bill — you're already paying for them.
A real run, in 14 seconds
Plan. Implement. Verify. Ship. Each phase runs in parallel across your reviewers — and they have to agree before the next one starts. (Three reviewers shown; bring as many as you like.)
Strengthen auth middleware · 12 files · 3 reviewers approved · click Merge in GitHub when you’re ready
Why Chorus
Same model writing and reviewing has the same blind spots. The fix is making different models grade each other.
Different training, different blind spots. Claude misses what Codex sees. Codex misses what Gemini flags. Run them together; the gaps shrink.
Claude Pro? ChatGPT Plus? Gemini? Chorus drives whatever you've authenticated. We never see your prompts and we never bill you.
Chorus is a binary on your machine. Not a SaaS, not a sidecar. Your code never leaves the disk it lives on. Nothing to sign in to. Nothing to leak.
Personas
A persona is a worldview prompt chorus prepends to the reviewer. The same Claude run as Sentinel hunts injection holes; run as Cartographer it flags Windows path separators; run as Translator it reads your error messages like a confused user. One model, ten roles — pick the one this PR needs.
Hunts secrets, injection, broken auth, supply-chain risk.
Catches when a change fights the existing pattern instead of joining it.
Windows / macOS / Linux assumptions, path separators, line endings, encoding.
N+1 queries, big-O cliffs, cold-path costs, latency budgets.
Reads labels, errors, empty states, help text — for normal users, not engineers.
"Who pays for this, when, and is there bill shock?"
Time-to-first-success: install path, error legibility, docs alignment.
Every new dep — maintenance, license, transitive footprint, install scripts.
What's not tested but should be — and what's tested but worthless.
Reads README + marketing copy alongside the diff and flags every lie.
voices × personas × phases — pair any model with any worldview at any step.
Five voices × ten personas × four phases = two hundred distinct reviewer configurations. Mix yours per template; fork ours; build your own. Personas are just markdown files in ~/.chorus/personas/ — you can write a new one in a minute.
Get started
Install the binary. Init detects your CLIs and registers MCP. Start brings up the daemon and cockpit. That's it.
$ npm install -g chorus-codes 🎉 Chorus installed! Two more commands to get going: ➡️ chorus init register MCP, seed templates, detect CLIs ➡️ chorus start bring up the daemon + cockpit on :5050
One npm command. No account, no signup, no cloud handshake. Chorus is a CLI that runs entirely on your laptop.
$ chorus init ▶ Initializing Chorus… ✓ database ready at ~/.chorus/chorus.db ✓ seeded templates: architect-review, bug-diagnose, code-review, red-green Detecting orchestrators… ✓ Claude Code MCP · auto-approved ✓ Codex CLI MCP registered ✓ Gemini CLI MCP registered ✓ OpenCode MCP registered ✓ Kimi CLI MCP registered ✓ Antigravity CLI MCP registered ✓ Grok Build MCP registered ○ Cursor not detected ○ Windsurf not detected ✓ Chorus initialized
chorus init finds the CLIs already on your machine and registers Chorus as an MCP server in each one. No keys to paste — Chorus rides on subscriptions you already pay for. Missing CLIs stay greyed out; install them later and re-run init.
$ chorus start ✓ Chorus started daemon PID 1234 Cockpit http://127.0.0.1:5050
chorus start boots a local daemon and the cockpit UI. Same engine, three doors: type chorus from the terminal, call it from inside Claude Code/Cursor via MCP, or open localhost:5050 in your browser. Pick whichever fits the moment.
Integrations
Sign in once with the CLI you already use. Add API keys for the rest. Three reviewers in parallel cost the same as one — because Chorus runs on your subscription, not ours.
Other multi-LLM tools route through their own paid APIs, so three reviewers means three times the bill. Chorus drives the CLIs you've already signed into, on the flat-rate plans you already have. The fourth reviewer costs nothing extra. So does the fifth.
What's inside
Two industry-standard multi-agent patterns. Two we packaged into a primitive. All four ship in the box — no glue code, no prompt-engineering homework.
All reviewers run at once against the same input. Three opinions in the time of one. A final pass distills them into a single verdict.
The author is one model family. The reviewers are different families. Shared blind spots don't survive contact with a different brain.
If they don't agree, the next round starts with the disagreement on the table. Capped at maxRounds — converge or stop. Never spin forever.
The implementer never sees the tests. The verifier runs them and reports failures in plain English. The model can't game what it can't read.
Pricing
Bring your own CLIs and you'll never see a bill from us. Top up later if you want a model you don't already subscribe to.
Drives the CLIs you've already signed into. Yours forever.
Gap-filler, not upsell. For models without a consumer CLI to subscribe to.
Open source under Apache 2.0. Cloud-hosted Pro and Teams tiers ship in v1.5 and v2.0.
FAQ
Straight answers. Still wondering about something? Open an issue on GitHub.
No. Chorus is a local tool. The daemon runs on your machine and talks to whatever CLIs you've already authenticated (Claude Code, Codex, Gemini, Kimi, OpenCode). On the free tier, your code, your prompts, and the model responses never touch our servers.
If you later choose to use bundled credits (v1.0+), only the prompts you explicitly route through credits go through our edge proxy — and even then, the proxy is stateless and never logs payloads.
Most existing tools are CLI-only orchestration engines. Chorus is the polished product layer on top — a web UI, MCP server, cross-CLI driver-agnostic design, threshold-aware quorum, and info-asymmetric red-green flows. Same agent patterns underneath; very different surface for the user.
If you live in your terminal and don't want a UI, the bare-metal alternatives are great. Chorus is for people who also want a panel they can share with a teammate.
Yes. Bring-your-own-CLIs is free forever. Paid options (bundled credits, hosted Pro/Teams) are purely additive — you never lose access to what's free today. The CLI orchestrator is also Apache 2.0 open source, so even if we vanished tomorrow, you could keep running it.
Yes — Chorus exposes an MCP server, so any MCP-speaking tool can call it. Cursor, Claude Code, Windsurf, and any future client that speaks MCP. You call mm.create_chat from inside your editor and the synthesis comes back as a tool result.
Two safeguards by default: (1) Chorus only writes to your filesystem during phases that are explicitly marked as "doer" — never during reviews or plans. (2) Every doer phase shows a diff and waits for approval unless you flip auto-approve on. If you turn auto-approve on, you've opted out of those gates.
Always commit before running. Chorus is open source, AS IS, no warranty. Like any AI tool, it can produce wrong or even destructive code. We're not liable for data loss — read the Terms for the full disclaimer.
Yes. The CLI engine and daemon are Apache 2.0 — patent grant included, corporate-friendly, fork-friendly. Repo: github.com/chorus-codes/chorus.
The hosted SaaS layer (multi-tenant, billing, premium templates, marketplace) is proprietary — but you never need to use it.
Next ship
v0.8 is here — install it now and you're set. Subscribe for v1.0: We send it once. No drip campaign, no newsletter.
No spam. No reselling. One-click unsubscribe — but you only get the one email.