Relay

Multi-agent coding, on your machine

Your agents. One workflow.

Relay runs the coding agents you already use (Claude Code, Codex, and models on your own GPU) as one team. One plans. Several build in parallel. Your tests run. Another reviews, and sends it back until it is right.

See a recorded run View on GitHub
The Relay dashboard showing a finished run: a local planner, three builders working in parallel (two local models and Codex), tests passing, and a local reviewer approving. The right side lists 33 changed files and the checks.
The real dashboard, showing the run recorded below.

How it works

  1. Plan

    A planner reads the repository and writes a brief and a file list for each builder, plus the reviewer's checklist.

  2. Build in parallel

    Builders work at the same time, each on its own files, in an isolated git worktree. Your checkout is never touched until you merge.

  3. Test

    Your own test command runs on the result. A failure goes straight back to the builders.

  4. Review

    A reviewer checks the change against the checklist. It approves, or the work goes round again.

Local-first

Agents run on your computer, in your repositories. Runs, transcripts and logs stay there. No telemetry.

Cloud and local models together

Mix Claude Code and Codex, signed in with your own subscriptions, with models in Ollama or LM Studio on your GPU.

You stay in charge

Watch every step live. Pause, step in, approve or reject at a checkpoint, and merge only what you want.

Recorded run Recorded, not live

A real run

Time
Outcome
Tests
Files

Timeline

    Files the builders wrote

      Run it at home, drive it from anywhere

      The PC with the GPU and the signed-in tools does the work. A laptop or phone only steers it.

      Your laptopPick a model, start a task, watch it, approve
      Private networkTailscale, or Cloudflare Access
      Signed gatePaired devices only, per-device permissions, audit log
      Home PCPlanner, builders, reviewer, git, tests, local models