Tap

What LeonTing1010/tap MCP server does

LeonTing1010/tap MCP server turns an instructed browser task into a saved flow plan that can be replayed deterministically. The plan uses a closed operation vocabulary and can define selectors, fetched fields, expectations, and a return transformation. Capture accepts a URL, an intent written in natural language, and optionally a site/name pair. With the pair, the flow is stored under ~/.tap/flows/<site>/<name>.flow.json; without it, capture returns a preview.

Saved flows are exposed as MCP resources using tap://{site}/{name} URIs. An agent can inspect available resources, read the argument schema and provenance, and then call run with matching arguments. A successful run includes both the raw observe data and the transformed return value. The raw observation is useful when auditing vehicle health or checking fields that the flow's return expression omits.

How it works

Capture performs the one-time AI-assisted inspection and produces the flow. Later calls dispatch the recorded operations rather than asking an LLM to reconsider the page. Runs use the user's authenticated Chrome session when the browser extension is configured, so cookies and credentials remain on that machine. Public pages and open APIs can instead use the Playwright runtime with an isolated profile.

The verify tool runs only the observe phase and reports live, drifted, or unreachable from operation outcomes, status codes, expectation predicates, and parsing results. It does not execute the act phase, although fetch operations still contact the live site. Per-operation CEL op.expect predicates can assert expected shapes or values. If an act operation may have committed but the process stopped before the result was known, mark records the confirmed outcome as committed or aborted.

Setup and configuration

The documented MCP stdio configuration runs the CLI with npx -y @taprun/cli mcp stdio. The CLI can also write host configuration with npx -y @taprun/cli embed <target>, including targets for Cursor and Claude Desktop. Claude Desktop additionally supports downloading and opening the tap.mcpb package.

For logged-in sites, the Chrome extension and native-messaging bridge are needed to reuse the live browser session. The setup flow registers the bridge and opens the extension page; installing the extension from Chrome is the remaining manual trust step. Public pages do not require the extension. The repository also documents Homebrew, a shell installer, platform binaries, and a --no-extension Playwright mode.

Tools and capabilities

• capture: Create a preview or persist a flow from a URL and intent.
• run: Replay a saved flow with zero runtime LLM tokens.
• verify: Check observe-phase health without performing the act phase.
• mark: Resolve an intent_uncertain record after checking the real outcome.
• MCP resources: Discover saved taps and read their argument schemas and provenance.

Limitations and notes

LeonTing1010/tap MCP server depends on the target site's current behavior and can return tap_drifted when selectors, responses, or expectations no longer match. Re-capturing with the same site and name overwrites the saved flow and is the documented repair path. A verification call is not entirely offline or strictly read-only because it probes live fetch targets.

Run failures use typed outcomes such as tap_not_found, tap_invalid, tap_aborted, credential_missing, arg_invalid, intent_uncertain, and runtime_unavailable. When a failure supplies a recovery call, the caller should issue that call; otherwise, the issue must be escalated to the user. The Chrome extension is specifically required for reuse of logged-in browser state, while isolated Playwright execution does not provide that live session.