Give your agent a prior-art tool

Patent77 hosts a Model Context Protocol server over HTTP at https://www.patent77.com/api/mcp. Three tools, the same engine as the API, and an anonymous trial mode, an agent can run its first real search before anyone creates an account.

Add it to your client

Claude Code

claude mcp add --transport http patent77 https://www.patent77.com/api/mcp

Add a key later with --header "Authorization: Bearer p77_live_...".

Cursor (.cursor/mcp.json)

{
  "mcpServers": {
    "patent77": {
      "url": "https://www.patent77.com/api/mcp",
      "headers": { "Authorization": "Bearer p77_live_..." }
    }
  }
}

Any MCP client (generic HTTP entry)

{
  "type": "http",
  "url": "https://www.patent77.com/api/mcp",
  "headers": { "Authorization": "Bearer p77_live_..." }
}

Anonymous vs. keyed: with no Authorization header the server runs in trial mode, 4 calls/min, 20/day per IP, search results truncated to the top 3 references, and reports unavailable. A paid-plan key (from settings after signup) applies your plan quota and full results.

The three tools

Served by tools/list from the same catalogue this page reads, so they can't drift.

search_prior_art

Run a prior-art search from a plain-English invention description - returns scored references (publication number, one-line overlap, elements hit) and a headline knockout risk.

Example arguments

{ "description": "A bicycle helmet with a built-in crash-detection accelerometer that texts an emergency contact" }

get_patent

Look up one patent by publication number - title, abstract, assignee, publication date, claims and the canonical source URL.

Example arguments

{ "pub_number": "US10123456B2" }

assess_patentability

Build the full patentability report - verdict, examiner-style narrative, element-by-reference matrix, limitations. Pass a prior search_id or a description for search + report in one call. Requires a paid key.

Example arguments

{ "search_id": "SEARCH_ID" }

Input schemas are returned verbatim by the server’s tools/list; the REST equivalents live in the API reference. Note that search_prior_art runs a real engine pass and can take a couple of minutes, set your client’s tool timeout accordingly.

A sensible agent workflow

  1. search_prior_art with the invention description, to get scored references and the knockout risk.

  2. get_patent on the top one or two references, to read the actual claims before drawing conclusions.

  3. assess_patentability with the search_id when the person asks for the full read. It needs a paid key, and the report has its own limits section your agent should pass on, not hide.

More in the agent onboarding guide and the machine-readable /llms.txt.