Reticle

What Reticle MCP server does

Reticle MCP server gives an AI coding agent a way to verify changes inside a running web or desktop application. A development-only SDK is added to the app, while the MCP server exposes tools for looking at the app, performing actions, observing resulting behavior, and asserting expected outcomes.

Verification returns one of three results: yes, no, or unknown. A successful check includes evidence such as the request caused by an action, resulting state changes, console output, and the source location associated with an element or failure. Unknown means the available evidence was insufficient; it is not treated as a pass.

The SDK can inspect network responses, application state from supported state-management libraries, console errors, request counts, and React renders. It supports web frameworks including React, Next.js, Vue, Svelte, and Angular, as well as desktop applications built with Electron or Tauri.

How it works

The agent interacts with the live application through a local bridge. Reticle drives a Chromium browser for browser-based applications and reads telemetry from inside the app instead of relying on screenshots. The bridge listens on 127.0.0.1:4400 and uses a local token stored under the user’s home directory.

A journey can be saved as a flow after it has been checked. Later, a single call can replay that flow deterministically without another AI model. The reticle gate command can fail a change when code associated with a saved flow has changed without a new passing run.

An optional Harness on the platform can describe a person and journey, drive the application, verify each step, and save successful results. A page HUD shows actions, verdicts, saved flows, and pinned notes while an agent works.

Setup and configuration

The repository provides an npm package named @reticlehq/server. Agent configuration uses one MCP entry that runs npx @reticlehq/server mcp. Project initialization can be previewed with npx @reticlehq/server init --dry-run; --no-mcp skips agent registration, and --files-only writes project files without continuing.

The SDK is intended for development builds. Production builds remove it, and a runtime guard prevents production connections. The setup may modify development dependencies, build configuration, a reticle-dev file, MCP configuration, and optional agent instruction files. Initialization appends to supported instruction files rather than overwriting them.

Tools and capabilities

Reticle MCP server supports these verification activities:

• Inspect the running application and its internal evidence.
• Perform UI actions and observe their network, state, and console effects.
• Assert expected results and receive a yes, no, or unknown verdict.
• Save and replay deterministic verification flows.
• Inspect source-file and line references connected to findings.
• Show live activity and results in the in-page HUD.

The server does not provide an arbitrary JavaScript evaluation tool. It also does not bypass client permission prompts. Captured passwords, tokens, and card numbers are replaced with [REDACTED].

Limitations and notes

The SDK must be present in the development version of the application to inspect internal behavior. Reticle does not include it in production builds. By default, anonymous command and tool usage counts are sent; reticle telemetry disable, RETICLE_TELEMETRY=0, or DO_NOT_TRACK=1 disables telemetry and other optional outbound calls.

Application data stays local unless the project is linked with reticle connect or RETICLE_API_KEY is used in CI. Linking enables configurable synchronization for runs, memory, and flows. The platform receives Harness steps when that feature drives an app, but secret field values remain local.

The server, CLI, SDK, adapters, core, and engine do not all use the same license: the SDK-related components are Apache-2.0, while the server and CLI use FSL-1.1-ALv2 with Apache-2.0 availability after two years.