claude-code-council

Stop letting one AI grade its own homework.

/council is a Claude Code skill. It runs the same question through two other models, Gemini and GPT, as local CLIs, then has Claude compare their answers and give a verdict. Under the hood it's gemini -p plus codex exec, with Claude doing the framing and the synthesis. The dressed-up version: Claude chairs a panel and the panel has no stake in Claude being right.

The panel calls run through CLIs you already log into (gemini and codex), so they don't spend any extra API keys — they draw on your existing Gemini and Codex CLI sessions. (Claude still spends its own tokens framing the prompt and synthesizing the replies; see What it actually costs.)

You:     /council is this migration plan safe?

Claude:  [asks Gemini and GPT in parallel...]

         ✅ Both agree: the down-migration is missing a backfill step.
         ⚠️ Disagreement: Gemini wants a feature flag; GPT says ship behind a
            transaction. (GPT) is right here because your DB supports DDL in tx.
         Verdict (Claude — chair): add the backfill, wrap in a transaction,
            skip the flag. Here's the corrected plan...

Real panel output is longer and messier than this. Models ramble, sometimes disagree on the wrong thing, and occasionally one refuses. The chairman's job is to cut through that — the value is in the reconciliation, not in any single reply.

---

Why

A model is wrong most often where it can't see its own blind spots. A second model that had no part in the first answer is the cheapest way to catch that. This skill makes it a one-command habit instead of three browser tabs and copy-paste.

• No extra API keys. It calls the gemini and codex CLIs you've already authenticated, not a paid API endpoint.
• No browser automation, no scraping, no token extraction. It only shells out to the CLIs you installed. (See Responsible use.)
• Graceful degradation. Only have one of the two CLIs? The council still convenes with one panelist.

It's a thin wrapper, and it's honest about that. The value isn't clever infrastructure — it's making "go ask two other models and reconcile the answers" a single command you'll actually use.

---

Install

1. Install at least one panel CLI

Panel	CLI	Get it
Gemini	gemini	gemini-cli, then run gemini once to log in
GPT	codex	OpenAI Codex CLI (the 2025 agentic CLI, not the retired 2021 Codex model), then run codex login

You need one or both. The skill works with whichever are present. If a CLI isn't installed or logged in, that seat is simply empty and the council runs with the rest.

2. Drop the skill into Claude Code

git clone https://github.com/<you>/claude-code-council.git
cp -r claude-code-council/skills/council ~/.claude/skills/council
cp -r claude-code-council/bin           ~/.claude/skills/council/bin
chmod +x ~/.claude/skills/council/bin/*.sh

The skill resolves bin/ask.sh relative to its own directory, so keep bin/ next to SKILL.md.

3. Verify

~/.claude/skills/council/bin/doctor.sh --smoke

You should see a ✓ and a live OK from each installed panel.

---

Usage

In Claude Code, just talk to it:

• /council review this function for race conditions
• council — is this architecture decision sound?
• get a second opinion on this approach
• have Gemini and GPT stress-test this plan

Claude (the chairman) will:

1. Frame one clear prompt and send the identical question to both panels in parallel.
2. Collect their independent answers.
3. Synthesize: agreements → disagreements → its own verdict, attributing every claim to (Gemini), (GPT), or (Claude — chair).

For code/design questions, paste the relevant diff or snippet into the prompt so the panels judge the actual code, not a paraphrase (scoped, and with secrets stripped — see the skill's redaction rule). For hard, multi-faceted calls you can also give each panel a different lens (Gemini → architecture/maintainability, GPT → edge cases/bugs) instead of the same prompt, so two passes cover more ground than one repeated. The same-prompt default stays best for a clean agree/disagree signal.

Configuration

All optional — override via environment variables:

Var	Default	Purpose
COUNCIL_GEMINI_BIN	gemini	path/name of the Gemini CLI
COUNCIL_GPT_BIN	codex	path/name of the GPT CLI
COUNCIL_GEMINI_MODEL	—	passed to gemini --model
COUNCIL_GPT_TIMEOUT	180	seconds before a GPT call is abandoned

---

How it works

                 ┌─────────────────────────┐
   your prompt → │   Claude  (chairman)    │
                 └───────────┬─────────────┘
                  same prompt │ in parallel
              ┌───────────────┴───────────────┐
              ▼                               ▼
       bin/ask.sh gemini              bin/ask.sh gpt
        → gemini -p                   → codex exec -s read-only
              │                               │
              └───────────────┬───────────────┘
                              ▼
                 ┌─────────────────────────┐
                 │ Claude synthesizes:     │
                 │  • agreements           │
                 │  • disagreements        │
                 │  • its own verdict      │
                 └─────────────────────────┘

bin/ask.sh is a thin, portable wrapper. It finds each CLI via which (or an override env var), calls it non-interactively, and returns clean text. That's the whole job.

---

What gets sent

This matters most when you're reviewing real code, so here it is plainly:

• A panel sees exactly the one prompt the chairman passes to ask.sh — nothing more. There's no automatic crawling of your repo, no file attachment, no hidden context bundle.
• If Claude decides a code snippet is needed for the review, it puts that snippet in the prompt text — so it's visible in the same command you'd see in Claude Code. Nothing is sent that isn't in that prompt string.
• codex is an agentic CLI that can read files in its working directory. To stop it reaching your repo, ask.sh runs it from an empty temp directory (created per call, deleted after). It literally has nothing to read but the prompt. This is enforced in code (bin/ask.sh, the gpt branch), not just promised here. It also passes -s read-only, codex's own sandbox flag for preventing writes and command execution (the exact guarantees are codex's to define and may vary by version).
• Whatever is in the prompt leaves your machine and goes to Google / OpenAI under their terms. Don't put secrets, credentials, or confidential data in a council question. The skill instructions enforce this, but treat it as your call.

If you need a model to review proprietary code you can't send to third parties, this skill is the wrong tool — keep that review inside Claude alone.

What it actually costs

The honest accounting, because "free" gets thrown around too easily:

• Panel calls: no Claude tokens, no separate API key. They draw on your existing gemini and codex CLI sessions, and are subject to those providers' quotas, rate limits, and terms. Hammer them in a loop and you'll hit limits.
• Claude's own tokens: still spent. Claude writes the prompt and then reads both panel replies back into its context to synthesize — so a council call costs more Claude tokens than just asking Claude directly. You're trading tokens for a cross-check, not getting one for free.

The pitch isn't "free AI." It's "a second and third opinion without juggling tabs or wiring up paid API keys."

---

Responsible use

This project does not bypass provider limits, scrape private endpoints, or extract OAuth tokens. It only shells out to locally installed, officially authenticated CLI tools that you set up yourself. Availability, quotas, and terms of service are controlled by each provider (Google, OpenAI) and may change at any time.

On attack surface: this skill adds none beyond the CLIs you already installed and logged into. It never grants them extra access. In fact it runs codex more narrowly than a normal session — -s read-only (codex's sandbox flag for preventing writes and execution) inside a throwaway empty directory. A panel returns text for the chairman to read; it isn't given a path to act on your repo. If you don't trust the gemini or codex CLIs to run on your host at all, that's a decision to make about those tools themselves — this wrapper only ever tightens their sandbox, never loosens it.

• Use it for personal and experimental workflows.
• Review each provider's terms before relying on it for production or team use.
• Never send secrets, credentials, or confidential data to a panel — anything you ask leaves your machine and goes to a third-party model. The skill's instructions enforce this, but it's ultimately on you.

---

Limitations (read before you rely on it)

No pretending here — these are real and you should know them going in:

• It rides on third-party CLIs. ask.sh calls the official non-interactive modes (gemini -p, codex exec), not screen-scraping of a spinner — so it's more stable than wrapping an interactive TUI. But these are still external tools on their own release schedule. A flag rename in a future gemini or codex can break a panel until ask.sh gets a one-line patch. doctor.sh --smoke tells you in two seconds whether each panel still works.
• The chairman is still one model. Claude frames the question and arbitrates the answers, so a badly framed prompt or an over-trusted panelist can still mislead it. The council widens the blind spot; it doesn't abolish it.
• Quotas are real. Panel calls spend your Gemini/Codex subscription quota. Run it in a tight loop and you'll get rate-limited. It's built for deliberate "check this decision" moments, not bulk automation.
• Not for code you can't share. Anything in the prompt goes to Google/OpenAI. For truly proprietary review, keep it inside Claude alone.

Found a broken flag or a sharp edge? PRs and issues welcome.

---

Prior art

The "convene other models for a second opinion" idea isn't new, and this didn't invent it — tools like the-council and ai-pair explore the same space, and the diff/snippet-in-context and per-panel-lens ideas were sharpened by looking at them. What this one leans on: no API keys (it rides your already-authenticated gemini/codex CLIs), an enforced privacy boundary for the codex panel (empty temp dir, in code), and a chairman that synthesizes rather than vote-counts. If those don't matter to you, the tools above may fit better — pick what works.

---

License

MIT — see LICENSE.

Not affiliated with Anthropic, Google, or OpenAI. "Claude", "Gemini", and "GPT" are trademarks of their respective owners.