CL
Claude Codeproposal
Extract the validation logic first. It reduces duplication and makes the behavior easier to test.
Open Source
Public Alpha
Multi-agent AI coding, without the manual relay.
You stay in control.
Codebate is an open-source orchestration workspace for Claude Code, Codex, and Cursor to discuss tasks, challenge plans, perform multi-agent code review, and safely execute changes before you approve them.
›_
npx codebate
Runs locally. Uses the coding-agent subscriptions and CLIs you already have.
CX
Codexreview
Agreed, but the error branch still leaks an internal message. Add a regression test before execution.
CR
Cursorconverged
The revised plan covers both concerns. No unresolved item remains in this round.
src/auth/validate.tsProposed change
Works with
Claude Code
Codex CLI
Cursor CLI experimental
The problem
Stop manually relaying answers between agents.
One agent proposes. Another finds a flaw. You copy the response back, repeat the context, and try to remember what was actually resolved.
“Take Claude’s plan, ask Codex to review it, send the objections back, then repeat until the plan is solid.”
Codebate turns that improvised loop into a structured local session.
Instead of this manual loop
1Copy the same project context into another tool
2Relay objections and answers between agents
3Lose track of agreements, open questions, and the latest plan
Codebate keeps one session, one task, and one official outcome.
How it works
From competing proposals to a decision you can trust.
Agents can disagree. Codebate keeps the discussion structured and makes the unresolved parts visible.
Discuss and challenge
Multiple coding agents see the same task and project context, then challenge each other’s assumptions and proposals.
Track real convergence
Agreements, open items, required user decisions, and malformed agent controls are handled explicitly — not hidden in a polished summary.
Execute, review, decide
Changes can be made in an isolated clone, reviewed against the exact tree, then accepted or discarded by you.
AI coding use cases
Practical guides for AI coding agents and agent orchestration.
Codebate focuses on a specific AI workflow: independent coding agents discussing the same task and producing a reviewable outcome.
AI coding agents
Learn how coding agents plan, review, and execute work — and where multi-agent coordination helps.
Read the guide →Multi-agent code review
Use independent coding agents to challenge a change and review the exact same code tree.
Read the guide →Claude Code, Codex & Cursor together
Replace manual copy-paste between coding tools with one structured collaboration session.
Read the guide →
Built for control
Your agents can collaborate without taking over your repo.
Codebate is designed around local execution, explicit approval, and honest failure states.
Runs locally
Your project and session stay on your machine. Codebate orchestrates the installed CLIs you choose.
Disposable execution clone
Execution happens outside your working repository, inside a temporary clone that can be reviewed safely.
Approval stays with you
The finalizer explains the official outcome. It does not silently rewrite it, and nothing is accepted without your decision.
Public Alpha
Use it on a real project. Tell us where it breaks.
Codebate is ready for early testing, not polished promises. The most useful feedback now is what happens in real repositories and real workflows.
- Was the multi-agent discussion better than using one agent alone?
- Did the agents actually resolve disagreements?
- Was setup clear on your operating system?
- What blocked you from trusting or accepting the result?
Product preview
FAQ
Questions about Codebate and AI coding agents
What is Codebate?
Codebate is an open-source local workspace that coordinates multiple AI coding agents in one structured session. Agents can discuss a task, challenge proposals, track unresolved items, review changes, and converge before you approve the final result.
Is Codebate an AI model?
No. Codebate does not replace Claude Code, Codex, or Cursor and does not provide model access. It orchestrates the supported coding-agent CLIs and subscriptions already installed on your machine.
Can Codebate perform multi-agent code review?
Yes. Codebate can let one agent execute a change in a disposable clone while other agents review the exact resulting tree. You decide whether to accept or discard the reviewed change.
Does Codebate modify my real repository automatically?
No. Execution is designed to happen in an isolated clone, and the reviewed change is only applied after an explicit user decision.
Which coding agents are supported?
Codebate supports Claude Code and Codex CLI, with an experimental Cursor CLI integration. Provider capabilities and safety restrictions can differ by operating system.
Does Codebate run locally?
Yes. Codebate runs on your machine and invokes the installed provider CLIs you select. Review the security documentation before using it with sensitive repositories.
Start a session
Let the agents debate. Keep the final call.
Install nothing globally. Open your project, run one command, and choose the providers you already use.