gavel 👨⚖️
Multi-model fusion for Claude Code.
Ask claude + codex + gemini in parallel, then judge their answers into one and act on it.
/gavel:fuse <task> asks three models the same thing: the Claude Code model you're running, OpenAI Codex, and Google Gemini. Codex and Gemini are read-only advisors; Claude is panelist #3 and the actor — the three answers are judged and synthesized into a single fused answer, which is then acted on. Only Claude writes to your workspace.
It runs the models through their local CLIs (codex, gemini), reusing your existing logins. No API keys to wire up, no MCP servers, no background jobs.
Inspiration
Gavel is inspired by OpenRouter's Fusion beats Frontier: dispatch a prompt to a panel of models, then have a judge synthesize their answers into one response that beats any single frontier model. Gavel brings that pattern into Claude Code — Codex and Gemini answer as advisors, and their answers are judged and fused into one before it acts.
Benchmark chart from OpenRouter's "Fusion beats Frontier" announcement (© OpenRouter), included for reference and attribution.
Install
In Claude Code:
/plugin marketplace add junkim100/gavel
/plugin install gavel@gavel
(Reload/restart if prompted.)
Local development — clone the repo and point the marketplace at the clone instead:
git clone https://github.com/junkim100/gavel.git
/plugin marketplace add /path/to/gavel # the cloned directory
/plugin install gavel@gavel
Setup
/gavel:setup
Reports whether codex and gemini are installed, authenticated, and recent enough, and offers to install whatever is missing. Authentication:
- Codex —
!codex login (install: npm install -g @openai/codex).
- Gemini — run
!gemini once to log in (OAuth), or export GEMINI_API_KEY=… (install: npm install -g @google/gemini-cli).
Gavel needs at least one advisor usable, but works best with both.
Commands
| Command | What it does |
|---|
/gavel:fuse <task> | Ask Claude + Codex + Gemini in parallel, synthesize one fused answer, then act on it. |
/gavel:ask <codex|gemini> <prompt> | Send a prompt to a single model and show its answer verbatim (no fusing, no edits). |
/gavel:setup | Check/install/auth the Codex and Gemini CLIs. |
/gavel:config [show | set <key> <value> | unset <key>] | View or change settings (model, timeout, panel) in the user or --project config file. |
How advisors stay read-only
Only Claude modifies your workspace. The two advisors are constrained differently because their CLIs differ:
- Codex runs in your project under its OS read-only sandbox (
-s read-only) — a hard boundary: it reads your code but cannot change it.
- Gemini has no equivalent read-only sandbox (its
plan mode doesn't stop shell-based writes), so gavel runs it isolated: in a throwaway directory with PWD/OLDPWD scrubbed, so it can't discover your repo path or make relative writes into it. It answers from the task text — include any code Gemini should see directly in your task. Note this is isolation, not a hardened sandbox: Gemini still inherits $HOME and could act on an absolute path you hand it, so don't paste untrusted content into a fuse expecting confinement.
- Prompts are passed via a temp file and reach each CLI on stdin, never through the shell or process arguments (so quotes /
$(...) / secrets in a task can't inject or leak).
Configuration
Defaults: Codex gpt-5.5-pro, Gemini gemini-3.1-pro, per-model timeout 1800s (30 min). These are the preferred defaults — if your account can't use them, gavel automatically falls back to whatever model the codex/gemini CLI itself defaults to (a model you explicitly set is always respected, never swapped). Override via env vars (GAVEL_CODEX_MODEL, GAVEL_GEMINI_MODEL, GAVEL_TIMEOUT) or a settings file — ~/.gavel/config.json (user) or ./.gavel.json (project).
Easiest way to change settings is the config command (no hand-editing JSON):
/gavel:config show # effective settings + which file each comes from
/gavel:config set timeout 600 # 10-min timeout, for all projects (~/.gavel/config.json)
/gavel:config set codex.model gpt-5.5 # pin a model (opts that provider out of auto-fallback)
/gavel:config set gemini.model gemini-2.5-pro --project # this repo only (./.gavel.json)
/gavel:config unset codex.model # restore the preferred default + auto-fallback
