ApiTap
The CLI, MCP server, and npm library that turns any website into an API — no docs, no SDK, no browser.
ApiTap is a CLI, MCP server, and npm library that lets AI agents (and you) browse the web through APIs instead of browsers. It ships with 6,400+ pre-mapped endpoints across 280+ APIs (Stripe, GitHub, Twilio, Slack, Spotify, and more) — ready to query on install. For sites not in the database, it captures API traffic from any website, generates reusable "skill files," and replays them directly with fetch(). No DOM, no selectors, no flaky waits. Token costs drop 20-100x compared to browser automation.
The web was built for human eyes; ApiTap makes it native to machines.
# Import 280+ APIs instantly — no browser needed
apitap import --from apis-guru --limit 100
Done: 87 imported, 3 failed, 10 skipped
1,847 endpoints added across 87 APIs
# Replay any imported endpoint immediately
apitap replay api.stripe.com get-listcharges limit=5
# Or capture a site's private API
apitap capture https://polymarket.com
apitap replay gamma-api.polymarket.com get-events
# Or read content directly
apitap read https://en.wikipedia.org/wiki/Node.js
✓ Wikipedia decoder: ~127 tokens (vs ~4,900 raw HTML)
No scraping. No browser. Just the API.
How It Works
ApiTap has three ways to build its API knowledge:
- Import (instant) — Import OpenAPI/Swagger specs from the APIs.guru directory of 2,500+ public APIs, or from any spec URL/file. Endpoints get a confidence score based on spec quality. No browser needed.
- Capture (30 seconds) — Launch a browser, visit a site, browse normally. ApiTap intercepts all network traffic via CDP, filters noise, and generates a skill file. Or use
apitap attach to capture from your already-running Chrome.
- Discover (automatic) — ApiTap auto-detects frameworks (WordPress, Next.js, Shopify) and probes for OpenAPI specs at common paths. Works without a browser.
All three paths produce the same artifact: a skill file — a portable JSON map of an API's endpoints, stored at ~/.apitap/skills/.
Import: OpenAPI spec → Converter → Merge → skill.json (confidence 0.6-0.85)
Capture: Browser → CDP listener → Filter → Skill Generator → skill.json (confidence 1.0)
Attach: Running Chrome → CDP attach → Filter → skill.json (confidence 0.8-1.0)
Replay: Agent → Replay Engine (skill.json) → fetch() → API → JSON response
↑ no browser in this path
Confidence Model
Every endpoint tracks how it was discovered:
| Source | Confidence | Meaning |
|--------|-----------|---------|
| Captured with response body | 1.0 | Full capture — response shape verified |
| OpenAPI import, high quality | 0.85 | Spec has response examples |
| CDP skeleton (real traffic, no body) | 0.8 | Endpoint exists, body was evicted from Chrome buffer |
| OpenAPI import, base | 0.6 | Thin spec, no examples |
Imported endpoints auto-upgrade to confidence 1.0 on first successful replay. The merge is additive — captured data is never overwritten by imports, imports fill gaps that capture missed.
Install
npm install -g @apitap/core
Claude Code — one command to wire it up:
claude mcp add -s user apitap -- apitap mcp
That's it. 12 MCP tools, ready to go. Requires Node.js 20+.
Note: npx @apitap/core mcp does not work due to npm scoped package bin resolution. Use npx apitap mcp or apitap-mcp instead.
Optional: To use capture and browse (which o
... View full README on GitHub
