Scry launches an MCP server for programmatic search over 164 billion web rows

What launched

A tool called Scry debuted on Hacker News this week with a pitch aimed squarely at AI agents: instead of searching for pages, query the underlying records. Scry is a Model Context Protocol (MCP) server that lets an agent run bounded, read-only programs across billions of internet documents, drawn from sources including Reddit, Hacker News, LessWrong, arXiv, Stack Exchange, Wikipedia and the prediction markets.

The figures on the project's site are live counters rather than marketing copy: at the time of the launch post it reported roughly 164 billion queryable rows across 43 sources, a held estate of about 419 billion rows (including sources under reviewed access that are not queryable by default), and around 35 billion rows newly indexed in the preceding 24 hours. The site says the counters are refreshed every five minutes.

How it works

According to the launch post, an agent first reads a live schema document, picks a source-native relation and any listed vector helpers, then writes a single query in Scry's SQL dialect for phrase matching, time windows, joins or vector composition. Every statement is read-only and runs under a deadline, a memory ceiling and a row cap on what returns, and result rows retain the source-native identifiers and provenance fields each relation provides. Recursive CTEs are supported; the site claims a 111-iteration branch-on-state loop returns in about 100 milliseconds.

The framing is deliberately set-based. A question such as every page that mentions one compound but not a certain patent, grouped by domain, since March, becomes a single query rather than something an agent assembles from repeated searches and scraped results.

Benchmark claim, with caveats

Scry's headline evidence is a run of Google's DeepSearchQA, a set of 900 research questions graded as exact sets. The site reports that an agent doing its research through Scry answered 71.8% of them fully correct, against 66.1% for Google's Gemini Deep Research Agent as listed on Google's Kaggle leaderboard, using the same judge model and grading prompt. Transcripts of every question, grade and tool call are published alongside. Two caveats worth holding onto: the site itself notes the corpus is still under half of Common Crawl, and this is a vendor-run comparison, so independent replication is the real test.

Freshness, access and pricing

Indexing lag is stated per relation in the schema rather than promised globally. Hacker News and LessWrong are indexed in under 15 minutes, while academic sources advance in layers as recorded folds re-run, with full-text extraction arriving on its own schedule.

Setup is a single MCP endpoint, mcp.scry.io, supported in ChatGPT (via developer mode and custom plugins), Claude, Claude Code, Codex, Cursor or any other MCP client. Agents without MCP support can use a plain HTTP API with a key from the project's dashboard. Sign-in creates the account, and nothing runs on the local machine.

Pricing is congestion-based, per the launch title, meaning cost tracks demand rather than sitting at a flat rate; the mechanics live on the site's pricing page. Scry describes itself as public-benefit infrastructure, subsidised for independent noncommercial research and carried by commercial engagements, with custom source builds and dataset work available by contact.

Why it matters

Most agent search today is pagination and scraping: fetch a page, extract, repeat. Scry is a bet that agents actually want relational access to the web's records, with joins, time windows, aggregates and provenance, delivered through MCP plumbing that ChatGPT, Claude and Cursor already speak. The congestion-pricing model is an unusual answer to the cost of querying at hundred-billion-row scale, and the published benchmark invites a direct comparison with Google's own research agents. If the numbers hold up under third-party testing, programmatic search becomes a credible alternative to the crawl-and-scrape loop. Even if they don't, it stands as one of the more ambitious attempts yet to treat the public web as a queryable database.