# Dispatch

Dispatch is an experimental, local-first CLI for comparing coding-agent harnesses on equivalent source states. It preserves execution and verification evidence, then lets a human evaluate the results.

- Canonical site: https://rundispatch.sh
- Core repository: https://github.com/dispatch-ai-org/dispatch
- Released real harnesses: Cursor Agent and Codex CLI
- Execution: local-first; source, diffs, logs, paths, and credentials stay local by default
- Verification: the same configured checks run against the baseline and completed candidate workspaces; checks do not automatically select a winner
- Cloud: optional, explicit opt-in evaluation contribution; enabling sync records consent and queues locally but does not upload
- Cloud status: sync is being enabled for early-access contributors; the hosted endpoint is not currently operational
- Automatic routing: not available yet; users choose harnesses for each run
- Current limitations: experimental v0.1 software, no hosted execution, dashboard, account system, billing, or automatic winner

## Start here

1. Install Dispatch and authenticate Cursor Agent and Codex CLI.
2. Run `dispatch init`, configure `checks.verify` in `dispatch.yml`, then run `dispatch doctor`.
3. Run both harnesses on one task, inspect the A/B evidence, and record a human evaluation.

## Public links

- Home: https://rundispatch.sh
- Download and install: https://rundispatch.sh/download
- Getting started: https://rundispatch.sh/docs
- Trust and sync boundaries: https://rundispatch.sh/trust
- Cloud direction and status: https://rundispatch.sh/cloud
- Human/agent sitemap: https://rundispatch.sh/sitemap.md
- Agent product guide: https://rundispatch.sh/AGENTS.md
- Latest Core release: https://github.com/dispatch-ai-org/dispatch/releases/latest