Coco Super Intelligence

495 personas across 13 departments, 229 skills[^installed] and 388 commands in the full catalog (coco 1.9.0, 2026-10-21).

Agent skills, slash commands and subagents for Claude Code, Cursor, Codex and other AI coding agents. One command installs them into the tool you already use. Coco's files and state stay on your machine. Your coding agent still uses its own model provider.

npx cocosuperintelligence

Open-core: the core is Apache-2.0, and Super Intelligence is proprietary. Relicensed from MIT to Apache-2.0 from 2026-11-05. Earlier releases remain under MIT. Files stay on your machine. No telemetry. A version check may contact GitHub once a day.

Why it exists

A hard call in a coding session usually gets one model's opinion, with no name on it and no record of who disagreed. CoCo seats a cross-team board of 495 personas, each modeled on the public writing of a named expert, answering from cited public stances, and returns one pick plus the dissent. They are not those people's views or endorsement. The same install carries the skills and commands that do the work after the verdict, and writes phase state to disk so a cleared chat does not erase the plan. It is not a new model and not a new harness. It installs into the one you already run.

See it

Personas are modeled on the public writing of named experts and are not those people's views or endorsement. See the persona disclaimer. To request removal or correction of a persona, see docs/personas/TAKEDOWN.md.

An illustration built from a real /SI-Decide run (Claude Sonnet 5.5, 30 September 2026, US$0.41). Persona names are replaced with role archetypes and the output is trimmed for length; no other words are changed. The full output, with the same replacements, is in docs/readme/terminal/si-decide-full.md. Note what it admits: the embedding router was offline, so it fell back to keywords, and stances applied beyond their source are marked extrapolated.

A real meta_select.py Stage-A run on this machine. The embedding router was offline (Connection refused), so this shows the keyword-fallback routing path (method: keyword, degraded: true), not embed cosine.

Quick start

Prerequisites: git, bash, Node.js 14 or newer (for npx), and python3 (it generates the /SI-* commands).

npx cocosuperintelligence installs the latest release. It clones the pinned release tag (currently v1.10.0, not floating main) into ./coco of the current directory, then runs install.sh. Run it from your home folder, so the clone is ~/coco and not a folder inside another project. On a fresh machine it took about 3 minutes, including the download.

npx cocosuperintelligence

Open Claude Code and run:

/SI-Decide "Should a 10-person startup move its public REST API to GraphQL this quarter?"

Newest main, which is ahead of that pin:

git clone https://github.com/coco-research/coco.git ~/coco && bash ~/coco/install.sh

Adapter flags and bundle selection: docs/install.md.

What it does

• Get one clear pick on a hard call, with the disagreement named and each stance cited.
• Ask in one line. /SI-Decide seats 16-32 of 495 personas across 13 departments. Add --debate for reaction rounds.
• Keep going after /clear. Decisions, phase state, and progress stay on disk.
• Install into Claude Code, Cursor, Codex, VS Code, or another adapter. Your model and your keys stay yours.
• Carry the verdict into the work: planning, review, verification, PRDs, and a pull request, with gates that have real exit codes.
• Stay local. CoCo adds no analytics. Set COCO_NO_UPDATE_CHECK=1 to skip the daily GitHub version check.

How it works

You type a question. The orchestrator scores 495 personas across 13 departments on domain fit, cell coverage, and productive conflict, then seats 16-32. Each seated persona is modeled on the public writing of a named expert, answers from cited public stances, and is not that person's views or endorsement. --debate adds reaction rounds before the tally. The decision is one pick, the dissents are named, and a stance check flags anything that does not trace to a source.

How the router scores teams: systems/superintelligence/README.md. What each editor actually receives: adapters/INDEX.md.

Status

Released on npm: npx cocosuperintelligence installs the latest release (CHANGELOG).

Measured on main, a default Claude Code install delivers 191 of the 229 skills (the rest are the opt-in security pack and Cursor-only skills), 388 commands and 34 agents (docs/delivered-counts.json).

Known gap: most GSD skills need the upstream GSD toolkit at ~/.claude/get-shit-done (see systems/gsd/README.md).

bash <(curl -fsSL https://raw.githubusercontent.com/coco-research/coco/main/bin/coco-bootstrap.sh)

CLONE="$(pwd)"
find ~/.claude ~/.cursor ~/.grok ~/.copilot -type l -lname "${CLONE}/*" -delete
for app in "Code" "Code - Insiders" "VSCodium" "VSCodium - Insiders"; do for base in "$HOME/Library/Application Support" "${XDG_CONFIG_HOME:-$HOME/.config}"; do [ -d "$base/$app/User" ] && find "$base/$app/User" -type l -lname "${CLONE}/*" -delete; done; done

The board

Super Intelligence summons a panel inside the session you already have open. Every contribution is from a persona modeled on the public writing of a named expert and grounded in a cited public source. It is not that person's views or endorsement.

/SI-Decide "Should we migrate our database to pgvector?"
/SI-AI-Decide "Which embedding model gives the best cost/performance?"
/SI-Eng-Pre-Mortem "Review our zero-downtime cache migration strategy"
/SI-GRC-Review "Check this customer onboarding flow for GDPR compliance"

495 expert personas, organized across 13 specialized departments:

• 342 generated slash commands. 17 cross-team commands (/SI-Decide, /SI-Tradeoff) plus 325 per-team commands (25 verbs × 13 teams, such as /SI-AI-Decide). Generated on your machine at install time. Nothing is uploaded.
• 98 expertise cells. Departments are split into cells so the router can seat a precise panel. The count is the sum of the cells fields in systems/superintelligence/registry.json.
• Two-stage routing. Stage A reads the meta-registry and picks the top teams. Stage B loads compact records for those teams only and seats 16-32 people with weights 0.40 domain, 0.30 cell coverage, 0.30 productive conflict, then pairs opposing views. If one team dominates, that team's own orchestrator takes the question.
• --debate. The panel argues in reacting rounds, then returns a verdict with named dissent instead of a false consensus.

The 495-persona roster was compiled with a systematic multi-tier workflow. Candidate generation was evaluated across local Qwen (via LM Studio), a hosted small model, and Gemini Flash, but Claude research agents proved the quality winner for resolving historical data and citing verifiable signal. Personas are checked by an advisory validation gate (validate_persona.py) that enforces five things: every required frontmatter field is present; at least 4 cited URLs resolve live (non-404); every public_stance carries an evidence_url, so no stance is uncited; at least 2 recent signals within 12 months (or 2 persistent signals for historical archetypes); and any pairs_well_with / productive_conflict_with slug refers to a real roster member. Two limits are worth stating plainly: the gate confirms that a stance is cited, not that a quotation is authentic, and home_team is only checked for presence, not validated against the list of real teams. A NEEDS-TOPUP verdict is advisory. It does not currently affect which personas a council selects at runtime.

Skills catalog

CoCo ships 229 skills (77 core + 152 across bundles)[^installed]. Each one is an instruction set with state, checks, and error handling, not a one-line prompt. A slice:

ui-ux-pro-max · frontend-design · brainstorming · writing-plans · test-driven-development · systematic-debugging · verification-before-completion · prd-generator · openai-agents · coco-loop · coco-ads · arch-index

System bundles

Six bundles install by default. --core-only skips them. --systems installs an explicit subset. reverse-skill is not one of the six: it vendors reverse-engineering and penetration-testing methodology under systems/reverse-skill/ and installs only with --systems reverse-skill. See systems/reverse-skill/README.md.

1. GSD. An orchestration engine of 68 skills and 24 agents for phased work on a large codebase. Most GSD skills also need the upstream GSD toolkit at ~/.claude/get-shit-done/, which this installer does not yet provide; see systems/gsd/README.md. Phase state lives in .planning/. Workstreams are isolated git checkouts.
2. Brain. A local knowledge-graph engine of 6 skills. SQLite indexes code entities, decisions, and project terms. The wiki generator writes a page per entity.
3. Super Intelligence. The 495-persona advisory board and its 342 generated /SI-* commands across 13 teams. This is the board above.
4. HyperFrames. 20 skills for video and motion, vendored from HeyGen's hyperframes under Apache-2.0.
5. M0. 4 skills for a cross-tool memory thread in one local SQLite file. Python standard library only.
6. Cognee. 3 skills for a local Cognee graph. They fall back to Brain when cognee server is not running.

/team is not a bundle. It is part of the core install (/team ship, /team plan, /team review, /team verify) and needs no flag. /team ship runs 6 development stages and 7 verification gates, and it can declare where files should land, then fail if they did not. The check is structural (paths exist), not a judgement that the code does what the component promised. Design adapted from lak7/devildev (Apache-2.0); no upstream code was copied. See CREDITS.md.

Technical specifications

A default Claude Code install delivers 191 skills, 388 commands and 34 agents, measured by running the installer. The repository holds 229 skills; the difference is the opt-in security pack and Cursor-only skills. The measured row is main.

_{Core install: 148 active assets (77 Skills, 46 Commands, 10 Agents, 15 Rules)
A plain install ships every bundle except the security-testing tooling, which stays opt-in via --systems reverse-skill.
Orchestration bundles: +68 GSD skills · +24 GSD agents · +20 HyperFrames skills · +13 Super Intelligence skills · +342 SI commands · +6 Brain skills · +4 M0 skills · +3 Cognee skills · 3 Workflows}

_{Core install ships 77 skills + 46 commands + 10 agents + 15 rules (148 active assets). A plain bash install.sh delivers 191 skills, 388 commands and 34 agents on Claude Code: every bundle in the default allow-list (GSD, Brain, Cognee, HyperFrames, Super Intelligence, M0) is installed by default. --core-only installs the core set alone, and --systems <list> installs an explicit subset — including reverse-skill, the security-testing bundle, which stays opt-in and is never part of the default set. Super Intelligence slash commands are generated locally at install time from the team registries — no command files are transmitted or stored remotely. adapters/INDEX.md records what each adapter actually delivers, measured by running it.}

FAQ

Credits

CoCo stands on other people's work. The short list:

• obra/superpowers (Jesse Vincent) for the engineering-discipline skills: brainstorming, debugging, TDD, plans, worktrees, review, verification, skill authoring.
• gsd-build/get-shit-done for the 68-skill / 24-agent GSD bundle.
• heygen-com/hyperframes (HeyGen, Inc., Apache-2.0) for the 20-skill video bundle under systems/hyperframes/.
• JCodesMore/ai-website-cloner-template for the structure behind clone-website.
• agents.md for the vendor-neutral context file the adapters follow.
• nickwinder/synthteam (Nick Winder) for the debate protocol behind --debate.
• latent-spaces/brag (Shunit Haviv) and HyperFrames (HeyGen) for the launch-video flow behind coco-ads.
• leonxlnx/taste-skill (Leon) for design-taste-frontend and redesign-existing-projects.
• HKUDS/CLI-Anything for the method behind cli-anything.
• Vercel: vercel-labs/web-interface-guidelines for web-design-guidelines, and vercel-labs/agent-skills for vercel-react-best-practices (link-only stub).

Full attributions, including audio (ende.app, Kenney.nl), are in CREDITS.md. If something influenced CoCo and is missing, that is a bug: open an issue.

Contributing

Coco is open-core: the core is Apache-2.0 and contributions are welcome. The Super Intelligence System (systems/superintelligence/) is proprietary. See systems/superintelligence/LICENSE. Relicensed from MIT to Apache-2.0 from 2026-11-05. Earlier releases remain under MIT.

• Start with CONTRIBUTING.md.
• Good first issues: help wanted / good first issue.
• Bugs and features: open an issue.
• Security: SECURITY.md. Conduct: CODE_OF_CONDUCT.md.

Releases are in CHANGELOG.md and on the GitHub Releases page.

Licence

Relicensed from MIT to Apache-2.0 from 2026-11-05. Earlier releases remain under MIT.

The Apache-2.0 licence in LICENSE covers the Coco core, which is everything in this repository except systems/superintelligence/ and the vendored third-party components below. The Super Intelligence System in systems/superintelligence/ is proprietary, licensed separately for your own use under systems/superintelligence/LICENSE, and the Apache-2.0 grant does not apply to it. The prior MIT text is kept at LICENSE-MIT-PRIOR. Vendored third-party components keep their upstream licences, which continue to govern them. They include HyperFrames in systems/hyperframes/ (Apache-2.0), reverse-skill in systems/reverse-skill/ (its upstream licence is documented in its README), and some skills under skills/. See CREDITS.md and each component's folder.

[^installed]: 191 of the 229 skills install by default on Claude Code, together with all 388 commands and 34 agents. The other 38 are the opt-in reverse-skill security pack (33, installed only with --systems reverse-skill) and 5 Cursor-only skills. Measured by running the installer, not by counting files: see docs/delivered-counts.json, and adapters/INDEX.md for every other IDE.

Coco Super Intelligence is built by Coco Research.