Reticle

What it does

Reticle lets a coding agent test its own change against the app it just edited. A small SDK goes into the app's development build. A bridge on your machine connects that SDK to the MCP server the agent talks to, so the agent can read the page, click around, follow network requests and console output, and ask whether an expected effect really took place.

Each check ends in a verdict. Yes is the only pass. No means it failed, unknown means the evidence could not decide, and no-fault means nothing was declared to check. The verdict comes with the request or log line behind it and the source file and line of the element involved. Journeys that passed can be saved as flows and replayed later, and `reticle gate` fails a CI run when a change touches a saved flow's code without a new passing run.

You can try it before wiring anything. Run `npx @reticlehq/server` with no arguments in the app's folder and it finds or starts the dev server, opens the page, follows up to 8 same-origin links and reports console errors and failed requests. It exits 0 only on a yes. Running `npx @reticlehq/server init` afterwards wires the app for real and registers the server with the agents it finds. It also puts hooks into each agent's project config, which keep every prompt so a verdict can be judged against what you asked for, and log the agent's own shell commands and edits under .reticle/. Staging and production builds can opt in to a production mode where each drive needs a ticket signed by a Reticle service and a policy refuses risky actions like payments or deletions.

The agent sees ten tools by default and calls the rest by name through `reticle_run`. Setting RETICLE_ADVERTISE_ALL_TOOLS does not show the whole registry, though. In tool-surface.ts it switches to a core set of 17 plus 11 extended tools, with the two meta tools on top, and everything else still goes through `reticle_run`.

What it needs

The app has to be wired once with `npx @reticlehq/server init` (add `--dry-run` to see the plan first). It adds the SDK (@reticlehq/react or @reticlehq/browser) and a build plugin as dev dependencies, edits the build config and writes agent instruction files and capture hooks. Add `--no-prompt-capture` to skip the hooks or `--no-mcp` to skip the agent files. Restart a dev server that was already running so it picks up the new config.

Limits

• Not covered Cannot see IndexedDB, Web Workers, closed shadow roots or cross-origin iframes
• Not covered Only partly catches races around a single action
• Platform Drives Chromium only, plus Electron and Tauri webviews, so there is no cross-browser run
• Not covered Meant for apps you own and run. Its README sends you to Playwright for outside sites
• Not covered Has no tool that runs arbitrary JavaScript in the page
• Limit Prompt redaction is a pattern match, so a secret that looks like an ordinary word can be kept in .reticle/
• Cost The Harness needs an account and credits beyond the free ones or the 14-day trial
• Limit Production mode works only with a Reticle service that issues signed tickets, and a linked API key
• Not covered The no-install check opens at most 8 same-origin links and stops at a sign-in page

Its own pages say it cannot yet see IndexedDB, Web Workers, closed shadow roots or cross-origin iframes, and it only partly catches races around a single action. The driven browser is Chromium (plus Electron and Tauri webviews), so there is no cross-browser matrix, and it is meant for apps you own and run locally rather than outside sites. It has no tool to run arbitrary JavaScript in the page. The Harness needs an account and runs on credits once the free ones or the 14-day trial are used up. Production mode needs a Reticle service that issues signed tickets and a linked API key.

Who it suits

Suits developers who let a coding agent change a web or desktop app they run locally and want proof from network, console and app state instead of a screenshot. Local use needs no account, and the first check runs without touching the project. Wiring it in changes the project (dev dependencies, build config, agent instruction files and hooks), so it fits poorly if you cannot alter the app or would rather keep agent prompts out of the repo folder. Its own README points to Playwright for sites you do not own or for testing across browsers, and to dedicated tools for pixel diffs, full accessibility audits and security scanning.