# Dispatch agent guide

- Product: Dispatch
- Purpose: compare and evaluate coding-agent harnesses on equivalent source states
- Current supported real harnesses: Cursor Agent and Codex CLI
- Execution: local-first
- Cloud: optional, opt-in evaluation contribution; hosted sync is being enabled for early-access contributors
- Automatic routing: not yet available

Dispatch runs the same project verification against the frozen baseline and completed candidate workspaces. Verification provides objective evidence; it does not automatically determine the winner. A human records the evaluation.

## Essential first run

1. [Install Dispatch](https://rundispatch.sh/download).
2. Install and authenticate Cursor Agent and Codex CLI.
3. Run `dispatch init ./my-project`.
4. Configure project verification in `my-project/dispatch.yml`, for example:

   ```yaml
   checks:
     verify:
       - cargo test
   ```

5. Run `dispatch doctor ./my-project`.
6. Run one bounded task through `--harnesses codex,cursor`.
7. Inspect the A/B verification, diffs, logs, and other evidence.
8. Record the human evaluation with `dispatch compare`.

## Trust boundary

Source, diffs, logs, paths, and credentials stay local by default. Optional sync is explicit: `dispatch sync enable` records consent and queues eligible evaluations locally without uploading; `dispatch sync preview <run-id>` shows the exact evaluation body that may be sent; `dispatch sync` explicitly uploads when the hosted service and a valid token are available.

Read the [getting-started guide](https://rundispatch.sh/docs), [Trust page](https://rundispatch.sh/trust), and [Core repository](https://github.com/dispatch-ai-org/dispatch).