Patent Search for AI Agents: API and MCP Server Guide

If you are building an agent that drafts patents, evaluates startup ideas, does technical due diligence, or scores inventions, it needs prior-art search as a tool, not as a suggestion to "check Google Patents." LLMs cannot answer novelty questions from their weights: the patent corpus is huge, precise, and updated weekly, and a hallucinated publication number is worse than no answer. Retrieval with verifiable citations is the only defensible architecture here, because a human (or a downstream attorney) must be able to click through from any claim the agent makes to the document that supports it. This page is the onboarding doc for wiring real, cited prior-art search into an agent.

What the API exposes

Patent77's engine (plain-English query in, professionally-structured search out, scored and cited references back) is available as a single-key JSON API and a hosted MCP server. Three endpoints do the work:

• POST /api/v1/searches — submit an invention description, get back ranked references with relevance scores, overlap summaries, and source links.
• GET /api/v1/patents/{pub_number} — fetch a specific document by publication number: bibliographic data, abstract, claims.
• POST /api/v1/reports — request a full patentability report: element-by-element mapping of your invention against the closest references, an examiner-style novelty and obviousness read, and a PDF. This is the paid deliverable.

Full request/response schemas are in the API reference; an OpenAPI spec is available at /api/openapi.

Try it in one curl, no key

Anonymous calls work out of the box, throttled to 4 requests/minute and 20/day, with results truncated to the top 3 references:

``bash curl -X POST https://patent77.com/api/v1/searches \ -H "Content-Type: application/json" \ -d '{ "query": "A collar-mounted ultrasonic device that deters dogs from barking using escalating tones triggered by an accelerometer detecting bark vibrations" }' ``

Authenticated calls use a single bearer key (no OAuth dance) and lift the truncation:

``bash curl -X POST https://patent77.com/api/v1/searches \ -H "Authorization: Bearer $PATENT77_API_KEY" \ -H "Idempotency-Key: search-9f2c1a" \ -H "Content-Type: application/json" \ -d '{"query": "..."}' ``

Notes that matter for agent builders:

• Idempotency keys. Pass an Idempotency-Key header on POSTs; retries with the same key return the original result instead of double-charging quota. Agents retry; design for it.
• Structured errors. Failures return machine-readable JSON error objects with stable codes (rate limits, quota exhaustion, malformed queries), so your agent can branch instead of parsing prose.
• Webhooks. Paid plans can register completion webhooks for async workflows, so long-running report generation does not block an agent loop.
• Machine-readable pricing. GET /api/v1/pricing returns current plans, quotas, and overage rates as JSON, so an agent (or a human) can reason about cost before committing.

The MCP server

If your stack speaks MCP (Claude, and most agent frameworks as of 2026), skip the HTTP plumbing entirely. The hosted server lives at https://patent77.com/api/mcp and exposes three tools:

Client configuration is one block:

``json { "mcpServers": { "patent77": { "url": "https://patent77.com/api/mcp", "headers": { "Authorization": "Bearer YOUR_API_KEY" } } } } ``

Anonymous MCP access is always on at the same trial limits (4/min, 20/day, top-3 references), so an agent can call the tool before anyone signs up for anything. Production quota requires a key from a paid plan.

Rate limits and pricing, plainly

Scale exists specifically for agent products: 600 req/min, 100 prior-art monitors, hosted MCP at production quota, and volume pricing beyond that. Current numbers always live at /pricing and /api/v1/pricing.

Design advice for prior-art tooling in agents

Five patterns we see work well:

1. Search early, report late. Have the agent run cheap searches during ideation or triage, and only trigger assess_patentability (the metered deliverable) when a human or a downstream step actually needs the full read.
2. Always surface citations. Every Patent77 reference carries its publication number and a link to the source document. Pipe those through to your end user untouched; a prior-art claim without a checkable citation is a liability, not a feature.
3. Write element-level queries. The engine performs best when the description names concrete technical elements ("capacitive soil-moisture probe recalibrated against a learned drying curve") rather than product marketing ("smart planter that keeps plants alive"). If your agent collects invention descriptions from users, have it decompose them into elements before calling search_prior_art; you will get materially better references.
4. Use monitors for long-lived subjects. Paid plans include prior-art monitoring: a saved invention is re-checked against newly published art, with webhook notification. For agents that track a portfolio of ideas over months, this is cheaper and more reliable than re-running searches on a timer.
5. Treat results as a screen, not an opinion. No automated search is exhaustive, and none of this is legal advice. The right agent behavior on a close call is to recommend a professional search or attorney review, not to render a verdict. (Context for calibration: human search firms charge $1,000-$3,000 per patentability search.)

If you are new to prior-art search itself, prior art explained covers what the corpus contains and what novelty and obviousness actually mean; it is worth encoding into your agent's prompt.

Sources and further reading

• Model Context Protocol
• USPTO Patent Public Search
• USPTO Open Data Portal
• Google Patents
• Espacenet (EPO worldwide patent search)