Conclave
A council of AI coding CLIs that brainstorm, converge on one plan, build, and cross-verify.
Orchestrates Claude Code, Codex, and Grok CLIs to plan together, split work, and adversarially check each other's output.
- Runs multiple coding CLIs on subscription logins, no API keys required
- Role-based permissions: planners/critics/verifier read-only, only the builder writes
- Adversarial cross-verification before output is accepted
- Shared memory layer for multi-round context
Conclave orchestrates several AI coding CLIs — Claude Code, Codex, and Grok — using their own subscription logins rather than raw API keys. The agents brainstorm a task in parallel, adversarially critique each other, converge on a single plan, split the build, and then cross-verify each other's output before anything is accepted.
A role-based permission map keeps the process safe: planners, critics, and the verifier run read-only, and only the designated builder can write to disk. A lightweight memory layer carries context between rounds so the council does not repeat itself. This is Milestone 1 of a broader multi-AI agentic-coding cockpit — the goal is code that survives review by three different models before it reaches you. Because the models disagree in useful ways, the convergence step surfaces trade-offs a single model would have quietly chosen for you, and the read-only critics keep that debate from ever writing bad code to disk.
How it works
- Task goes to Claude Code, Codex, and Grok CLIs
- Agents brainstorm in parallel, critique each other
- Council converges on a single plan
- Work is split; only the builder writes to disk
- Agents cross-verify output before acceptance