Conclave

Give each model a task. Keep the goal in one place.

Conclave

Orchestrate multi-agent workflows from a single CLI session. Conclave allocates effort by task and delegates across Claude, Codex, Gemini, and Grok. Route specialized tasks, right-size subagent models to the work, review results across LLMs, or find consensus through a multi-LLM council. Each job records what ran and what passed its checks, so the coordinator can integrate results toward the goal.

Built by Giovanni Carosso

Open ↗·Workflows·Source·License: Apache-2.0

How it works

I built Conclave to coordinate work across models from one session. The lead session defines the goal, delegates tasks through a local CLI, and integrates the results. Conclave handles routing, response review, bounded repair, and job records.

Route by the work required: Claude for complex reasoning and planning, Codex for implementation and tests in a local sandbox, Gemini for large-context analysis, and Grok for current web evidence. The coordinator combines the findings into one result, using review or council when useful. Focused briefs can reduce repeated context; parallel research can reduce waiting. Additional agents also consume tokens, so the useful comparison is cost and time per accepted result against a single capable model. The workflow guide covers these tradeoffs.

Each job retains its inputs, attempts, and checks. Default review evaluates returned text; repository changes still need build and test evidence.

Hypothetical workflows

Turn a technical demo into a working example. Gemini analyzes a recorded demo through an external video integration and returns timestamped observations. Grok checks the current API documentation. Claude turns the findings into an implementation plan, and Codex builds and tests the example. The coordinator combines the evidence, code, and test results.

Diagnose an ML pipeline regression. A script compares run metrics, and a smaller model summarizes the relevant logs. Codex reproduces the failure and tests a fix. If the evidence suggests an interaction between preprocessing, training, and evaluation, the coordinator brings in Claude for a broader analysis. The task expands only when the evidence calls for it.

Choose a research method. A smaller model extracts common fields from a set of papers. Gemini compares supplied methods sections when substantial context is needed, while Grok checks current software documentation. For a decision with competing assumptions, a Claude–Codex council can compare the tradeoffs. The coordinator checks the citations and produces one recommendation.

Model sizes and review settings are choices made by the coordinating session and routing policy. Start with the simplest adequate workflow and add calls when they address a specific gap.