patlas

npx add-mcp 'https://patlas.dev/mcp'

How to use it

1. Ask directly

Describe an idea in plain English to Claude, Cursor, or any MCP-compatible AI client. Patlas searches real USPTO/EPO patent data and has your AI reason over the actual claims text — ranked matches, not a guess.

2. Scan your own codebase

Connect Patlas to a codebase-aware AI agent (Claude Code, via the free patent-scan Skill) and it reads through your product's own source code to surface specific mechanisms and features worth checking — things you built but never thought to search for — then runs each one through real patent search automatically.

Your code never reaches Patlas. Your AI agent reads it on its side and sends Patlas only a short, plain-English description of each mechanism to search. (As with anything you do in that agent, your code is processed by the agent's own AI provider under its terms — not by Patlas.)

3. Turn a scan into a Word report for your attorney

Once a codebase scan finds something worth a closer look, ask your assistant to package the findings into a formatted "Patent Landscape & Prior-Art Audit" — a one-time $299 purchase that produces a Word (.docx) document you can hand to an attorney or investor instead of a chat transcript, and that reviewers can mark up with comments and tracked changes. It's delivered straight to your own signed-in session — never posted at a public link, because unfiled invention details shouldn't be. Every cited patent is independently re-verified before it appears in the report.

For teams

Team accounts

One flat $199/month bill covers up to 8 seats (100 calls/seat/month) — every team member signs in with their own GitHub or Google account, no shared credentials. Ask your assistant to "create a team called Acme Corp" to get started; anyone can start one regardless of their own current tier.

Patent monitoring (watch alerts)

Watch a technology area for new filings — Patlas checks once a day and posts to a webhook you supply when something new matching your query shows up. Available on Pay-as-you-go, Subscription, and Team (up to 5 watches per account).

Embed Patlas in your own product

A plain REST API (not MCP-specific) for approved integration partners white-labeling patent search under their own brand — admin-assigned, metered per call. Not self-serve: contact us to discuss an integration.

Why the output is grounded, not just fast

The risk with AI-generated patent analysis is hallucination — an LLM confidently inventing a patent ID or a claim quote that doesn't actually exist. Patlas validates every similarity claim and every quoted claim-language snippet against the real, retrieved patent text before it's returned. If a claim can't be grounded in the actual source document, it's dropped or corrected — not surfaced to you as fact. This is enforced in code, not just prompted. It reduces hallucination risk; it doesn't make the analysis complete or a substitute for professional review — see the disclaimer below.

What Patlas sees — and doesn't

Sees

• The invention description or question you type, when you call search_prior_art, get_patent_detail, or compare_claims — processed to generate the AI analysis, and used to query USPTO/EPO for patent data.
• If you use patent-scan against your own codebase: only the short, plain-English description of each mechanism your AI agent writes after reading your code — not the code itself. Generating a paid report from a scan sends that same mechanism/summary text plus the patent IDs you're citing — never your code — so the report can verify each citation and render the document.
• A numeric user ID from GitHub or Google, depending on which you sign in with. That ID is all we keep to recognize your account.
• Usage metadata for every call to a product/analysis tool: tool name, timestamp, token counts/cost estimate, and which patent IDs came back.
• Billing identifiers (Stripe customer/subscription ID) if you're on a paid tier — never your card number, which Stripe handles directly.

Doesn't see

• Your source code — during a patent-scan, your AI agent reads it and sends Patlas only plain-English mechanism descriptions. The code itself never reaches Patlas or whatever powers the analysis behind it. (Your agent's own AI provider processes it, as with any use of that agent.)
• Your plaintext API key — only a SHA-256 hash of it is ever stored, even by us.
• Your password — sign-in happens on GitHub's or Google's own page. Your email address or GitHub username is passed to us during sign-in but never stored.
• Your payment card details — Stripe handles those entirely.

Full detail in our Privacy Policy, including exactly which providers your data is sent to and why.

Pricing

Tier	Price	What you get
Free	$0	2 calls/month — 7 calls/month through November 30, 2026 (launch offer) — search_prior_art + get_patent_detail
Pay-as-you-go	$3/call	Unlimited calls, all tools including compare_claims and landscape_summary
Subscription	$49/month	20 calls/month, all tools
Team	$199/month	Up to 8 seats, 100 calls/seat/month, all tools shared across the team
Landscape & Prior-Art Audit report	$299 one-time	A Word (.docx) report from a codebase scan, delivered to your own session — not a subscription, buy one whenever you need it

No seats requirement, no annual contract, no sales call — Team's 8 seats are a cap, not a minimum.

Get started

Sign in with GitHub or Google — no separate signup, no API key to manage or lose.

No AI client? Use the web app

Don't have Claude Code, Claude Desktop, or Cursor set up? Use Patlas directly in your browser — same tools, no MCP client required. Sign in the same way in both places and it's the same account.

Claude Code

claude mcp add --transport http patlas https://patlas.dev/mcp

Then type /mcp in Claude Code and choose Authenticate. Or add it to your project's .mcp.json:

{
  "mcpServers": {
    "patlas": {
      "type": "http",
      "url": "https://patlas.dev/mcp"
    }
  }
}

Claude Desktop or claude.ai

Settings → Connectors → Add custom connector, and paste https://patlas.dev/mcp.

Other MCP clients

If your client supports remote HTTP servers, use https://patlas.dev/mcp directly. If it only supports local (stdio) servers, use the mcp-remote bridge:

{
  "mcpServers": {
    "patlas": {
      "command": "npx",
      "args": ["mcp-remote", "https://patlas.dev/mcp"]
    }
  }
}

The first time you use it, a browser window asks you to sign in with GitHub or Google.