Search skills/

Claude Council Claude skills

Claude Code plugin that sends one question to several AI models and CLIs in parallel, shows their answers side by side and has Claude synthesize them.

hex/claude-council 4 skills Code review & qualityPlanning, TDD & debugging
Our verdict Recommended with caveats

A well-built, tested multi-model council, but it needs keys or subscriptions for other vendors and sends your question and attached code to each of them, so read the data section first.

4/5 Depth expertise shipped Note Security Review before installing 3d Maintenance Updated 3 days ago 911 Adoption stars OK License MIT, commercial OK

What Claude Council does

Claude Council adds commands to Claude Code that send one question to several AI providers in parallel and show their answers side by side. You run /claude-council:ask with a question. The council-execution skill calls scripts/run-council.sh and saves the transcript under .claude/council-cache, a folder that ignores itself in git. It prints the answers verbatim, then Claude writes a synthesis of agreement, disagreement and a recommendation, and names the premise that could change the answer. Supported providers are Gemini, OpenAI, Grok, Perplexity, Kimi and OpenRouter over API keys, the codex, Antigravity, grok, kimi and cursor-agent CLIs on your existing subscriptions, and local ollama. If you have no provider, --local runs blind Claude subagents, and that skill tells Claude to say that agreement between same-model members is weak signal. The --agents flag (deep-execution) runs parallel Claude analysts that question each provider, ask follow-ups and return a schema-checked analysis with confidence ratings. Other options cover roles, debate rounds, auto-attached files, screenshots, background jobs, export and quiet mode. Opt-in extras include a live pane inside Claude Code, a heads-up side-check where one seat watches Claude's work, and a stop-gate review of the diff before Claude stops. Most of the logic lives in shell scripts with tests that run in CI. docs/REFERENCE.md lists every flag and what gets sent.

Good for

  • Architecture and framework decisions where one model's bias could mislead you
  • Debugging dead ends that need a second or third opinion
  • Developers who already pay for several model APIs or CLI subscriptions

Not for

  • Code you cannot send to external AI vendors
  • Users with no API keys or subscriptions who would only get a Claude-only panel

What's inside depth 4/5

4 skills averaging 5.7 KB of instructions on top of a large shell codebase (268 files, 132 executable), 52 test files run in CI, and long reference and architecture docs.

Skills
4
Reference docs
3
Scripts
0
Tests / evals
52
Templates
0
Avg SKILL.md
5.5 KB

Counted from the repository file tree. How we score depth.

Dependencies and costs

  • curl and jq
  • macOS, Linux, or Windows through Git Bash
  • At least one provider: an API key for OpenAI, Gemini, xAI, Perplexity, Kimi or OpenRouter, a logged-in codex, Antigravity, grok, kimi or cursor-agent CLI, or a local ollama model (otherwise only the Claude-only --local mode works)
  • CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1 for the optional pane
  • Telemetry: none found. The plugin does send your question, and any attached or auto-included file contents, to every provider you select. The opt-in heads-up side-check also sends recent turns, tool results and Claude's Write/Edit text to one seat (codex by default).

Security review

Review before installing

This plugin sends your questions to several AI providers (via their APIs with your own keys, or via agent CLIs you have installed) and shows their answers in a pane. It installs a Stop hook and a Claude Code hooks module. An optional heads-up side-check, off by default, sends recent turns, tool results and file edits to one seat (codex by default) on every check without asking each time. The plugin documents that the model-callable council tool asks for confirmation before sending a question.

Worth knowing

  • installs_hooksRegisters a Stop hook that runs scripts/stop-review-gate.sh and loads the council-pane hooks module (pane.tsx), which polls run directories and can launch the heads-up side-check.hooks/hooks.json:3
  • Makes network requestsSends council questions and status probes to provider APIs such as OpenAI, Gemini, xAI, Perplexity, Moonshot and OpenRouter.scripts/check-status.sh:140
  • Needs an API key for its own serviceEach provider's API key, such as OPENAI_API_KEY or GEMINI_API_KEY, is passed to curl through a mode-600 config file for that provider's own API.scripts/lib/retry.sh:12
  • Relies on a CLI or package we could not reviewSeats can run external agent CLIs (codex, antigravity, grok, kimi, cursor-agent, copilot, opencode, claude, ollama) whose code is not in the repo..claude-plugin/plugin.json:4
  • Changes files outside your projectThe maintainer screenshot script edits ~/.claude.json project trust and plugin store files, then restores them on exit. It also launches Claude Code with bypassPermissions.assets/capture/capture.py:179

What we checked

The repository (268 files) on Oct 29, 2026. Not a full review: we read 27 of 97 code files; it relies on code outside the repository: codex CLI (OpenAI), used as the default heads-up seat and as a council seat; antigravity (agy), grok, kimi, cursor-agent, GitHub Copilot CLI and opencode CLIs, used as optional seats; ollama local model server.

NVIDIA SkillSpector static scan: 323 pattern hits (Privilege Escalation 91, Prompt Injection 81, Tool Misuse 40, Data Exfiltration 35, Agent Snooping 27, analysis-evasion 24). Most hits are documentation mentions (CHANGELOG notes about keychains, Docker config and codex config.toml isolation), test fixtures (an 'rm -rf /' string in notices.test.ts) and cleanup of the plugin's own cache and watch directories. The bypassPermissions and --allow-dangerously-skip-permissions hits are in the maintainer's screenshot capture script and an e2e test, not in the user-facing commands. allowed-tools entries are limited to the plugin's own scripts plus wc and head.

A review, not a certification. Skills change after we check them — read the files before installing.

License

MIT, commercial use allowed.

Can it run unattended?

Not really. It answers one question per run and waits for you to read the synthesis; background jobs and the stop-gate review help during a session but there is no long-running loop.

Alternatives to Claude Council

Claude Octopus

nyldn/claude-octopus

An escalation layer that runs tasks past up to twelve external AI providers (Codex, Copilot, Ollama, Grok, more) with a consensus gate, plus 63 skills.

With caveats Check first Depth 4/5

delegate-skills

amElnagdy/delegate-skills

Twenty skills that hand a coding task to another agent CLI (Codex, Cursor, Aider, Claude and many more) in the background, then leave the diff review and commit to you.

Recommended Broad access Depth 4/5

Orca

stablyai/orca

An open-source desktop and mobile app for running several coding agents in parallel git worktrees, plus 11 thin skills that hand agents the app's CLI.

With caveats Check first Depth 2/5

FAQ

What is claude-council?

It is a Claude Code plugin by hex that asks several AI models the same question in parallel, shows each answer verbatim and has Claude write a synthesis of where they agree and disagree. It works with API providers, subscription CLIs such as codex and cursor-agent, local ollama, or Claude subagents as a fallback.

How do I install claude-council?

In Claude Code run /plugin marketplace add hex/claude-marketplace and then /plugin install claude-council. You need curl and jq, plus at least one provider: an exported API key such as OPENAI_API_KEY, a logged-in CLI like codex, or a local ollama model. Run /claude-council:status to check what it found.

Is claude-council safe to use on private code?

The code itself showed no hidden network calls or obfuscation, but by design it sends your question and any attached files to every provider you pick. Use ollama or --local if code must stay on your machine, and keep the opt-in heads-up side-check off unless you accept that one seat sees Claude's edits.

claude-council vs claude-octopus: what's the difference?

Both run multiple AI models against the same task to surface disagreements. claude-council is a Claude Code plugin with a side-by-side pane, debate and role modes, a Claude-only local fallback and an experimental mode where codex specialists work on their own branch. Compare the provider lists and workflows of each before choosing.

Sources: GitHub repository (stars, license, file tree, last push Oct 26, 2026). Review based on the README and SKILL.md files as published.