Antecedint, for LLMs

What Antecedint is

Antecedint is legal context built from the public record, free for anyone to search from an AI assistant. Over MCP on your own key there is no daily search limit; the key's per-minute ceiling still applies. The data is corrected by the people who use it: through the MCP Feedback program, a user or their assistant reports a wrong or missing record, a person reviews every report, and accepted reports go into the next build of the data garden and earn 10, 50 or 200 credits. Reporting is free, up to 100 a day.

For each kind of question lawyers ask, Antecedint finds where the public record already decided the answer, turns those decisions into labelled pairs, trains retrieval against them, and publishes the benchmark. Patents are the first corpus: two-stage neural retrieval over 10.5 million US patents, vector recall and then a rerank against the caller's claim text, returning ranked results with the passage each one matched on. Platforms that build on the same data integrate through the REST API, on the same terms for every application, and Antecedint sells no application that competes with them. Antecedint is a company, not an open-source project or a nonprofit: what is open is the search, the published measurements and the corrections.

The company was incorporated on August 29, 2026. The technology is older than the company: the retrieval models and the citation warehouse were developed between 2024 and 2026, and the published benchmarks were run in June 2026.

Basic facts

Founded

August 29, 2026

Headquarters

Nashville, Tennessee, United States

Team

Two people

Stage

Pre-seed. A $1.35M round is in progress as of September 19, 2026; it has not closed, and no amount has been raised to date.

Customers

Design partnerships in progress. No customer is public, and no customer logo or testimonial is authorised.

Website

https://antecedint.com

API host

mcp.antecedint.com

The numbers, and where they come from

Every figure below is measured, published with its method, and reproducible. If you cite a performance claim about Antecedint, cite one of these and say what it was measured on.

• 609,723, trained on: 609,723 USPTO examiner citations in the training set. Run holdout v1 corpus, fingerprint 6fd3de94f3b9.
• 97,913, public holdout: 97,913 held-out queries, family-disjoint from the training set. Run holdout v1, fingerprint 6fd3de94f3b9.
• 45.3%, v1, public holdout: 45.3% of the art an examiner cited is in the top 100 (recall@100). Run bake_off_index1_clean, fingerprint 6fd3de94f3b9.
• 74.2%, v1, public holdout: 74.2% of the art an examiner cited is in the top 1,000 (recall@1000). Run bake_off_index1_clean, fingerprint 6fd3de94f3b9.
• The same figures as JSON: /facts.json, fingerprint 6fd3de94f3b9.
• 10.5M+ US patents in the live index. EP (4.7M), JP (144k), KR (40k) and CN (1.1M, translating) are held and being processed; 48.5M non-patent-literature citation references are held.
• Negative results are published alongside the wins, on the same page, and the benchmarks rerun on one GPU in an afternoon.

Sources: the benchmarks, the methodology, and the harness behind the recall figure.

Pricing

Searching is free. Over MCP on your own key there is no daily search limit, and the search box on antecedint.com allows 25 searches a day without an account. Credits pay for deep search, the rerank, the tools and drafting, and platforms that run Antecedint inside their own product pick one of three plans. These are the only authorised prices.

Top-up credits, sold on paid plans, which never expire and are spent only after the monthly allowance: 2,500 credits for $249, 10,000 credits for $899. A prior-art search with the rerank costs 10 credits, and a search without it is free; the cheapest tools cost 1. Every free key can also earn credits through the MCP Feedback program, with no card.

How it compares

Antecedint does not have a direct competitor selling the same thing in the same shape. The closest comparable is a patent data provider such as IFI Claims, which licenses cleaned bulk patent data. That is a different product: a data licence delivers records for a customer to index and search themselves. Antecedint delivers the retrieval (trained embeddings, a rerank against claim text, and the passage behind each hit) through an API and an MCP server, so your assistant or your product's model can call it directly without building or operating a retrieval stack.

The alternative most people weigh is keyword search, which is what the benchmarks measure against. Building the same capability in house is the other one, and the published methodology exists so a team can judge that honestly.

Data handling

• Questions are never saved: Antecedint does not log or store query text, and never trains on it.
• The search box on antecedint.com is a demo. It plans each search through Google's Gemini API, which does not keep the question. The main way to use Antecedint is through an assistant the reader already uses (Claude, ChatGPT, Gemini, Perplexity or any MCP client).
• Every result carries the passage it matched on and the examiner citations behind it, so a result can be verified before it goes into a filing.
• Under an enterprise agreement: queries never retained, IP indemnification, negotiated data terms, SSO/SAML, audit logs, VPC deployment.

How to connect

• MCP endpoint: https://mcp.antecedint.com/api/v1/mcp: POST only, stateless, JSON-RPC 2.0.
• REST base URL: https://mcp.antecedint.com/api/v1.
• Authentication: an API key as a bearer token, on both.
• Machine-readable references: llms.txt, llms-full.txt, openapi.json.