← all repositories
coco-research/coco

495 fake experts walk into your coding session

CoCo layers a routed advisory board, 227 skills, and persistent state on top of Claude Code, Cursor, or Codex — without running a model of its own.

coco
Not currently ranked — collecting fresh signals.
star history

What it does

CoCo is a portable layer of Markdown and YAML that installs into the agent harness you already use and upgrades it with three things: an “advisory board” of 495 named expert personas across 13 departments, a large library of skills and generated slash commands, and disk-persistent project state that survives /clear. It doesn’t run a model or its own agent loop — your host harness does the actual work; CoCo supplies the orchestration structure.

The interesting bit

The persona selection is a genuinely staged algorithm: a cheap registry scan picks 1–4 relevant teams, then a scoped pass loads compact persona records and greedily assembles a 16–32-person panel with deliberate conflict pairing (security vs. growth, say), so you get named dissent instead of false consensus. The README also honestly labels its own dogfood demo as running in degraded keyword-fallback mode — a rare touch of candor.

Key highlights

  • 495 personas in 13 departments, from Engineering (70) to Sales/GTM (23), subdivided into 70 expertise cells for precise panel assembly
  • 342 generated slash commands (/SI-Decide, /SI-Eng-Pre-Mortem), built locally at install time from team registries
  • Every persona stance must carry a live, non-404 evidence_url; a validation gate enforces frontmatter, citations, and recent signals
  • /team:ship runs a 6-stage build pipeline with 7 verification gates and atomic git commits per step
  • 100% local, no telemetry, vendor-neutral across Claude Code, Cursor, and Codex

Caveats

  • Open-core: the MIT core is open, but “Super Intelligence” — the headline feature — is proprietary, own-use only
  • The validation gate confirms a stance is cited, not that a quotation is authentic; NEEDS-TOPUP verdicts don’t actually affect runtime selection
  • The 495-expert framing is prompt engineering over your existing model, not access to real people

Verdict

Worth a look if you want structured deliberation and persistent state bolted onto your existing coding agent. Skip it if you’re allergic to persona theater or need a fully MIT-licensed stack.

Frequently asked

What is coco-research/coco?
CoCo layers a routed advisory board, 227 skills, and persistent state on top of Claude Code, Cursor, or Codex — without running a model of its own.
Is coco open source?
Yes — coco-research/coco is open source, released under the Apache-2.0 license.
What language is coco written in?
coco-research/coco is primarily written in Python.
How popular is coco?
coco-research/coco has 630 stars on GitHub.
Where can I find coco?
coco-research/coco is on GitHub at https://github.com/coco-research/coco.

heatdrop uses Google Analytics to see which pages get read — nothing else. Your call. How we handle data.