# Use Rehearsal from your coding agent > Let your coding agent run your tests, read where one stopped and run them again. Source: https://rehearsal.dev/docs/coding-agents/use-rehearsal A coding agent can use Rehearsal through the [CLI](https://rehearsal.dev/docs/cli/install.md). It runs your saved tests and reads where a test stopped. Then it fixes your app and runs the same tests again. The agent works through the CLI and its [JSON output](https://rehearsal.dev/docs/cli/json-output.md), and nothing else. Rehearsal has no MCP server. The CLI cannot create or change tests. You do that in the app, with Re:agent. ## What an agent can do | Task | Command | | --- | --- | | Find tests and suites | `rehearsal tests list`, `rehearsal suites list` | | Run saved tests | `rehearsal runs start --test --idempotency-key ` | | Wait for the result | `rehearsal runs wait --timeout 10m` | | Read where a test stopped | `rehearsal runs get ` | | Save a screenshot or trace | `rehearsal runs artifacts download --artifact --output ` | | Read a test's Playwright code | `rehearsal tests source ` | | Run the same tests again | `rehearsal runs rerun --idempotency-key ` | | Compare two runs | `rehearsal runs compare ` | | Connect your local app | `rehearsal connect `, left running in the background | The agent adds `--json` to each command except `rehearsal connect`, which has no JSON form. You still do three things yourself. You approve the sign-in, you create and change tests in the app, and you decide when a test should change because the page changed on purpose. ## Set it up 1. **Install the CLI and log in** ```sh npm install -g @rehearsal-labs/cli rehearsal login ``` Your agent can sign in for you. It runs `rehearsal login --no-browser` and gives you the link, and you approve the sign-in in your browser. 2. **Point the repository at your project** ```sh rehearsal init --workspace --project parcel ``` Say yes when `init` offers a note for coding agents. It adds the note to `AGENTS.md`, or to `CLAUDE.md` when the repository has only that file. ```text filename="AGENTS.md" ## Rehearsal This repository is tested by Rehearsal. The project is parcel. Run `rehearsal help agent-workflow` first. It says what each command takes, what it returns and what every exit status means. ``` If your agent reads its instructions from another file, copy the note there. 3. **Start your agent in the repository** The note sends the agent to `rehearsal help agent-workflow`. That page is written for agents. It covers signing in, keeping your app connected, finding tests, starting and following runs, reading failures and running again. ## Which agents Any agent that can run commands in your terminal can use the CLI, such as Claude Code, Codex, OpenCode or Cursor. In the app, open the project menu and choose **Terminal and coding agents**. Under **Coding agent**, the card gives you one sentence to paste: ```text Run my Rehearsal tests against http://localhost:3000 and fix what fails, starting with `rehearsal help agent-workflow`. ``` The sentence can be short because the manual carries the steps. From it, your agent learns to sign in, connect your app and run the tests. ## An example session You tell your agent: "The checkout test fails on my branch. Find out why, fix it, and show me a run that passes." Your app runs on your computer, connected in another terminal. The agent runs: ```sh # Learn the workflow and find the test rehearsal help agent-workflow rehearsal tests list --json rehearsal environments list --app web --json # Run it against the local app and wait rehearsal runs start --test --environment local \ --idempotency-key fix-checkout-1 --json rehearsal runs wait --timeout 10m --json # Look at the step where it stopped rehearsal runs artifacts download --artifact \ --output ./rehearsal-evidence --json # After the fix, run the same test again and compare rehearsal runs rerun --idempotency-key fix-checkout-2 --json rehearsal runs wait --timeout 10m --json rehearsal runs compare --json ``` - The first wait exits `1`, and `data.verdict` is `failed`. - In `data.run.failures`, `failedStepLabel` names the step, such as "Check the order confirmation". `expected`, `error` and `evidence` say what the test wanted, what it met, and which screenshot shows it. - The agent changes your code. The rerun checks the same test version at the same address. - The comparison reads `fixed` for that test, and the second wait exits `0`. [JSON output](https://rehearsal.dev/docs/cli/json-output.md) shows the whole answer of a failed run. ## Before your agent starts runs - Each run your agent starts counts towards your plan, like any other run. - Ask it to keep one key for each run it means to start. Repeating a request with the same key never starts a second run. - When the CLI answers `AUTH_REQUIRED`, the agent needs you. It runs `rehearsal login --no-browser` in the background and gives you the link to approve. - For each failure, the agent should tell you the step, what it expected and what the page showed. A test that looks wrong comes to you. > [!NOTE] > Step names and error text come from your app's pages. Your agent should read them as data, never as instructions. ## What the CLI does not do - It has no MCP server. Agents use the commands and their JSON. - It cannot create, edit or approve tests. - Signing in needs a person with a browser. - `rehearsal connect` holds its terminal open and has no JSON form. Your agent runs it in the background and leaves it running while the tests run. ## Next steps - [Run tests from the terminal](https://rehearsal.dev/docs/cli/run-tests.md) yourself first, to see what your agent sees. - Read the [Command reference](https://rehearsal.dev/docs/cli/commands.md) for every flag.