ScryMCP Server

Scry inverts the usual shape of agent search. Instead of handing back ten links for a query string, it exposes a large archive of public internet documents as queryable relations and lets the agent write read-only SQL against them. The question an agent actually has — every post mentioning this compound but not that patent, grouped by domain, since March — becomes one statement rather than a loop of searches and scrapes.

The corpus spans 43 sources, among them Reddit, Hacker News, LessWrong, arXiv, Stack Exchange, Wikipedia and the prediction markets. The site publishes live counters: roughly 164 billion queryable rows at the time of writing, a held estate of about 419 billion, and on the order of 22 billion rows added in the preceding 24 hours across live capture and backfill.

How an agent uses it

The server exposes a schema call and a sql call. The convention is that the agent reads the schema first — which also states how far each relation's coverage extends and how fresh it is — then issues SQL with a literal LIMIT. Rows come back carrying their source ids, timestamps and URIs, so a claim can be traced to the document it came from rather than to a summary.

Turing-complete queries. Recursive SQL and fixpoint programs (the μ operator) let one statement do graph walks, negation and whole-population aggregates with conditional branching, rather than the agent driving iteration from outside.

Vector search. embeddings.* relations support semantic ranking and compositional embeddings, which can be combined with ordinary source-native SQL filters in the same query. A separate authenticated rerank endpoint will order documents you already hold.

Bounded by construction. Every statement is read-only and runs under a deadline, a memory ceiling and a cap on returned rows, so a badly-shaped agent query fails fast instead of running away with the budget.

Connecting

One URL — https://mcp.scry.io — serves every client, over stateless streamable HTTP with no session handshake. First connection sends you through an OAuth approval against your scry.io account, which binds that client to its own API key that can be budgeted or revoked from the dashboard. It is documented for ChatGPT developer mode, Claude.ai, Claude Desktop and Cowork, Claude Code, Codex and Cursor, and works with any MCP client. Without an MCP client there is a plain HTTP API: GET /v1/scry/context, GET /v1/scry/schema, then POST /v1/scry/query with a text/plain body and a bearer key.

Status and licensing

Scry is in open alpha and openly described by its author as a one-person project. Personal keys are licensed for individual, non-commercial research; commercial use requires a separate engagement. Treat coverage and the live row counts as moving numbers rather than guarantees.