Set the line: one agent per stage
Investigate with one model, implement with another, review with a third. Pick a stage, then plug it into a different harness. The config, the diagram above, and the terminal below follow along.
Stages
Harnesses
Select a stage, then a harness to repatch it.
Drive it from the keyboard
Everything happens in one terminal. This is the real layout with your line from above. Click it and press r.
| Harness | Status | Structured output | Used by |
|---|
Checked by model-router doctor. No model request was sent.
Unattended when it can be
Research, review, and the closing brief run in the background with their own timeouts. You're notified when a run needs you.
Interactive when it matters
Implementation opens the harness's own terminal. Detach with Ctrl+B, quit mr, and come back to the same screen.
Handoffs, not scrollback
Each stage passes a structured summary, verdict, findings, and test evidence to the next.
Interlocks you can't pull past
The router coordinates agents. It doesn't take their keys, widen their permissions, or touch your Git history without asking. Each one stays locked until you take an explicit step to release it.
No commits by default
Only a ship stage commits, on a new branch, after you confirm a preview.
Your logins stay yours
No stored credentials. Provider keys are read from your environment at launch and redacted in every record.
Native permissions
Each harness keeps its own approval rules unless a stage explicitly overrides them.
Trust before commands
A repo's config can't run checks or custom adapters until you run model-router trust.
No silent model swaps
A failed stage stops and offers retry, a replacement model, or resume.
Clean exits
Every harness process group ends with its stage. Optional bubblewrap or sandbox-exec per stage.
Harnesses on the line
Install and log in to the ones you use. model-router doctor checks what it finds without sending a request.
- Codex structured output
- Claude Code structured output
- Gemini
- Amp
- Cursor
- OpenCode
- Aider
- Goose
- Copilot
- Pi
- omp
- Your CLI custom adapter
# try the whole flow offline, no model calls cargo run -- --config examples/demo.toml # see exact commands, nothing launched model-router preview --task 'Respect Retry-After' # run it unattended in an isolated worktree model-router run --task-file spec.md --unattended --worktree # retry a stage with a different model model-router resume r-0418 --retry --model opus # read the brief and the diff model-router report r-0418 --diff # let other agents drive it model-router serve --mcp