conclave
Health Warn
- License — License: MIT
- Description — Repository has a description
- Active repo — Last push 0 days ago
- Low visibility — Only 7 GitHub stars
Code Fail
- fs module — File system access in .releaserc.json
- rm -rf — Recursive force deletion command in scripts/conclave-run.sh
Permissions Pass
- Permissions — No dangerous permissions requested
This CLI tool runs coding tasks and questions through multiple AI models (like Claude and Codex) simultaneously to provide comparative reviews, second opinions, and refined implementation plans. It acts as a local orchestrator for the AI CLIs already installed on your machine.
Security Assessment
The tool acts as a shell script wrapper around your existing AI CLIs, executing them locally and writing temporary files to disk. It does not make direct network requests to hidden servers, proxy your data, or request dangerous system permissions. There are no hardcoded secrets. However, the rule-based audit raised a critical flag: the source code contains an `rm -rf` recursive force deletion command within a shell script (`conclave-run.sh`). While potentially just used for routine cleanup of temporary files, this command can be dangerous if it handles paths incorrectly or is exploited. Overall risk is rated as Medium due to this destructive file system operation.
Quality Assessment
The project is very new and has low community visibility with only 7 GitHub stars, meaning it has not been extensively peer-reviewed. However, it is actively maintained (last push was today) and is properly licensed under the permissive MIT license. The developer README is clear and transparent about what the tool does and does not do.
Verdict
Use with caution — the orchestration concept is highly useful and runs entirely locally, but you should inspect the `rm -rf` command in the shell script before executing to ensure it will not accidentally delete important directories.
Multi-model AI collaboration for Claude Code
Consensus before commit.
Conclave is a local multi-model CLI for developer decisions. It runs your coding tasks through more than one model, using the CLIs you already have.
Compare when you want perspectives. Converge when you want them to argue.
Modes
| Mode | What it does | When to use it |
|---|---|---|
/review |
parallel critique on a diff | before commit or PR |
/consult |
parallel second opinions on a question | when stuck |
/converge |
one model drafts, another challenges, a coordinator tracks issues until they resolve or hit a cap | when the plan matters |
/review and /consult compare - same prompt, multiple models, consensus highlighted.
/converge deliberates - two roles (implementer and reviewer), one ledger, bounded rounds.
What it's not
- Not agent soup
- Not a hosted orchestration product
- Not an always-on framework
- Not "ask 5 models and pray"
Two roles, one ledger, bounded rounds.
There is no Conclave-hosted service and no Conclave proxy. Conclave invokes the CLIs already on your machine - Claude, Codex, Gemini, whatever you have - and runs locally. The underlying CLIs still call their providers' APIs; that's up to you to configure.
Quick start
git clone https://github.com/caiopizzol/conclave ~/dev/conclave
cd ~/dev/conclave
bun run register
Requires: bun, jq, and at least one AI CLI on your PATH (Claude, Codex, Gemini, ...).
/review and /consult work with any one or more configured CLIs. /converge needs the default config's claude-opus and codex keys both enabled, because the implementer and reviewer roles are hardcoded to those keys in the MVP. Swap support is planned.
First successful run
Stage a small diff, then from Claude Code:
/review
Each configured model critiques the diff in parallel. Consensus items (flagged by two or more models) are marked separately. Single-model findings are more likely noise.
For a question instead of a diff:
/consult "why is the table rendering broken?"
When the plan itself matters, not just the diff:
/converge "design a rate limiter for the auth endpoint"
/converge takes longer than the other two because it runs a review/revise loop. Expect 2-5 minutes. The session writes every prompt, artifact, and ledger snapshot to a temp directory so you can audit what the models actually said.
When to use what
- Fast signal before a commit:
/review - Quick second opinion on an approach:
/consult - A plan you want models to challenge and refine:
/converge
Advanced
The engine is a shell script that reads a config, runs tools in parallel, and returns JSON. Custom commands, custom scopes, and the engine's JSON contract are covered in docs/advanced.md.
The /converge coordinator has its own design document: docs/converge.md.
Why
More models does not mean better. Consensus is the signal.
- One model flags an issue → might be noise.
- Two or more models flag the same issue → probably real.
- One model drafts and another challenges until they agree → a plan you can actually trust.
Conclave does not replace your judgment. It structures the output of multiple models so disagreement becomes useful instead of overwhelming.
Inspired by LLM Council.
License
MIT
Reviews (0)
Sign in to leave a review.
Leave a reviewNo results found