browser-to-api

Browser to API

Replay-driven API discovery. Consume a browser-trace capture, pair its CDP request / response events, templatize observed URLs, infer JSON schemas from samples, and emit an OpenAPI 3.1 document plus a human-readable coverage report.

This skill does not capture traffic. It is purely offline post-processing on top of browser-trace's cdp/network/*.jsonl buckets. The two skills compose:

browser-trace    →  .o11y/<run>/cdp/network/{requests,responses}.jsonl
browser-to-api   →  .o11y/<run>/api-spec/index.html + openapi.yaml + client.mjs

When to use

• The user wants an OpenAPI document for a third-party or undocumented website API.
• The user has a browser-trace run and wants endpoints + schemas extracted from it.
• The user is building a client/SDK against a site that doesn't publish a spec.
• The user wants a coverage report showing which flows would broaden the spec.

If the user wants to capture traffic, send them to browser-trace first.

Two-step workflow

1. Capture with browser-trace (and optionally bodies via browse network on)

# Local example against an existing debuggable Chrome target
TARGET=9222
 
node ../browser-trace/scripts/start-capture.mjs "$TARGET" my-site
browse open

browse network on is optional but strongly recommended — without it, the spec has no response-body schemas (the CDP firehose used by browse cdp does not embed bodies). With it, both request bodies (already captured by CDP) and response bodies are joined into the trace by CDP requestId.

2. Generate the spec

discover.mjs auto-detects <run>/cdp/network/bodies/. To use a body capture from elsewhere (e.g. didn't snapshot, want the live browse network dir), pass --bodies <path> explicitly.

3. Open the HTML report

After discover.mjs finishes, always open the generated HTML report:

The report is a self-contained HTML file (no server needed) that shows each discovered operation as an expandable card with variables, client usage, request/response examples, and a generated client.mjs snippet at the bottom. This is the primary deliverable — always open it for the user.

CLI flags

Output layout

What you get from browse cdp and browse network

Two complementary capture sources:

discover.mjs will pull bodies from a browse network dir if you pass --bodies <path> (or stash them under <run>/cdp/network/bodies/, which is auto-detected). The matching is by requestId — browse network writes that into each request.json as id, and we join directly.

What changes when bodies are present:

• ✅ Path templating, query-param schemas, status codes, content-types — same either way.
• ✅ Request-body schemas — postData from CDP is enough; bodies dir is a nice-to-have for non-postData cases.
• ✅ Response-body schemas — fully inferred from real samples. Without bodies you get { description, content: <mimeType> } skeletons.

The report flags every endpoint that has no response-body sample.

Automatic noise filtering

The normalize stage automatically classifies and drops infrastructure noise:

• Tracking / analytics — paths containing /track, /pixel, /beacon, /impression, /pageview, /dag/v*
• Bot defense — Akamai (/akam/), fingerprint payloads (sensor_data), obfuscated multi-segment paths
• Session plumbing — /session, /authenticate/start, cookie consent, A/B experiment endpoints
• HTML page renders — GET requests returning text/html (the rendered page, not the API)

This typically drops 60-80% of captured traffic. The --include flag can rescue a false positive.

GraphQL / multiplexed endpoint decomposition

When a single endpoint (like /dapi/fe/gql) is called with different operationName values, the skill automatically splits it into separate logical operations. Each gets its own:

• OpenAPI path entry (e.g. /dapi/fe/gql [Autocomplete])
• Request/response schema inferred from only that operation's samples
• Curl example and variables table in the report

Detection works on body fields (operationName, method, action) and query params (opname, op). This covers GraphQL (APQ and inline), JSON-RPC, and similar dispatch patterns.

Limitations

• Coverage is bounded by the captured flow. Endpoints not exercised in the trace will not appear. The skill cannot prove completeness.
• Schemas are inductive, not contractual. A field might be optional on the server even if every sample contained it.
• Auth is observed, not specified. The skill records auth-shaped headers in an x-observed-auth extension but won't claim a security scheme.
• Path templating is heuristic. Numeric / UUID / hex / slug patterns are detected per segment. Ambiguous URLs are flagged in confidence.json.
• Redaction is best-effort. Default redactions cover common credentials, but app-specific secrets may slip through; use --redact for known custom headers/keys.

Best practices

1. Drive the flows you want documented. The richer the browser-trace, the richer the spec.
2. Use --origins for noisy sites. A marketing page hits dozens of analytics hosts; restrict to the API origin you care about.
3. Inspect report.md first. It has curl-ready examples and response samples for every discovered operation.
4. Bump --min-samples to 2+ when you want only confidently-shaped endpoints in the final doc — drop the long tail.
5. Pair with browse network on when response-body schemas matter. The CDP firehose alone has request bodies but not response bodies.

For pipeline internals and the file format reference, see REFERENCE.md.