TL;DR: Scry is an MCP server that gives an AI agent read-only SQL over a corpus of roughly 54 billion indexed rows: Reddit comments, Hacker News, arXiv papers, SEC filings, Wikipedia, prediction markets. You add one URL to Claude, ChatGPT, Codex or Cursor, and the agent writes bounded SQL against the archive instead of assembling a guess from ten search links.
What Scry does
Web search hands your agent ten links and a summary somebody else wrote. Scry starts somewhere else: a corpus with a published schema, and SQL as the query language. Ask for every post that mentions a compound but not a related patent, grouped by domain, since March, and you get a result set with source ids, timestamps and URIs attached to each row instead of a paragraph.
The mechanics are simple once you know them. Your agent calls the schema tool to read the live relation catalog, picks the relations that could hold the answer, and writes one query with a literal LIMIT. The service runs it and hands back rows plus accounting fields.
Two numbers describe the scale. The live wire endpoint at https://api.scry.io/v1/stats/wire reported 53,956,370,500 queryable rows on the morning I checked. The landing page claims it scans every byte of indexed Hacker News text in 0.86 seconds, which works out to 15.5 billion characters at 18 GB per second.
The tool roster goes past plain search:
schemafor the live relation catalog and per-relation coveragesqlfor bounded, read-only queriesembedandvectorsfor semantic search and mintingrerankfor reordering documents you supplymemory,feedbackandwhoamifor session state and connector health
Nothing runs on your machine. Every query hits their infrastructure, which is what makes the install a two-minute job.
Where the data comes from
| Family | What it holds | Example relation |
|---|---|---|
| Social | Reddit comments and posts, Hacker News, Stack Exchange, LessWrong, Bluesky, Mastodon | reddit.comments |
| Writing and web | Wikipedia, Substack posts and comments, mailing lists, a web crawl | wikipedia.articles |
| Scholarship | arXiv, PubMed, Crossref, OpenAlex citations, full-text papers, books, patents | academic.catalog |
| Markets | Kalshi and Polymarket | markets.catalog |
| Records | SEC EDGAR filings | sec.edgar_documents |
| Code | Package registry records | packages.catalog |
The source descriptions are unusually honest about their limits. The Reddit archive runs on pushshift, per the founder on the launch thread, and is described as near-complete from 2005 onward, with the caveat that archive months land with lag and the live tail captures only a fraction of current volume. The practical consequence matters: a zero in the results can mean absence in the indexed data rather than absence at the source. Each relation publishes its own extent and its known holes.
Getting it running
Claude Code takes one command, then an approval inside a session.
claude mcp add --transport http scry https://mcp.scry.io
Inside a running session, run /mcp and approve. Adding --scope user makes the server available in every project. For headless or CI runs, skip the browser and send a key as a header instead.
claude mcp add --transport http scry https://mcp.scry.io \
--header "Authorization: Bearer $SCRY_API_KEY"
Codex takes the same URL with its own command pair.
codex mcp add scry --url https://mcp.scry.io
codex mcp login scry
Any other MCP client takes a plain JSON block. Claude.ai uses Settings, then Connectors, then Add custom connector with the same URL. ChatGPT needs developer mode switched on under Security and login, then Plugins, the plus button, and the server URL.
{
"mcpServers": {
"scry": { "url": "https://mcp.scry.io" }
}
}
There is also a packaged skill that keeps query patterns and schema notes next to your agent.
npx --yes skills@latest add exopriors/skills
Writing the first query
The SQL dialect is not Postgres, but it is close. Token matching uses helper functions instead of LIKE, and every statement needs a literal LIMIT of at most 10,000 rows. Start at 20 and page with a WHERE clause on the ORDER BY key.
The first Hacker News item to mention bitcoin:
SELECT hn_id, original_author, original_timestamp, title
FROM hackernews.items
WHERE hasToken(search_text_lc, 'bitcoin')
ORDER BY original_timestamp ASC
LIMIT 5;
Where Reddit talked about it in 2013:
SELECT subreddit, count() AS n
FROM reddit.comments_popular
WHERE created_utc >= '2013-01-01'
AND created_utc < '2014-01-01'
AND hasToken(search_text_lc, 'bitcoin')
GROUP BY subreddit
ORDER BY n DESC
LIMIT 10;
Two habits make this work in practice. Have the agent call schema first, since the catalog is live and each relation carries its own predicates and holes. And pre-flight a wide statement for free with the x-scry-explain: 1 header, which runs index analysis instead of a full query.
The query surface also computes rather than just retrieves. WITH RECURSIVE runs on the same endpoint, and a JSON program body runs fixpoint graph walks across 36 registered edges, covering citation graphs, Twitter reply and follow graphs, and pivots into GitHub, Bluesky, YouTube and TikTok.
How the billing works
Queries are free while the machine has slack. The price endpoint reported a billing regime of free_slack with a congestion multiplier of zero at the time I checked, and the site says most questions cost a fraction of a cent to a few cents.
When load rises, a congestion multiplier gets posted and the price climbs with it. You consent to a ceiling in advance, and the query either runs at that price or refuses. New accounts start with a five dollar grant.
The unit of account is a second of query time. The x-scry-max-seconds header buys a time box of whole seconds, from 1 to 60, charged in full when the query is admitted and not refunded for the part you did not use.
| Knob | What it controls |
|---|---|
x-scry-max-seconds | Seconds of query time you buy, 1 to 60, default 1 |
x-scry-budget | Ceiling on raw machine burden, even when the query is free |
x-scry-max-multiplier | Highest congestion price you will accept before a refusal |
x-scry-explain: 1 | Free index pre-flight of a wide statement |
Idempotency-Key | Binds one statement so a retry reads settled status instead of running again |
A query can also be paid per call with x402, in USDC on Base, with no account at all. The posted rate was 0.05 USDC per second at par, plus a congestion premium.
Every response separates two numbers. burden_nanodollars is what the machine did, spend_nanodollars is what you paid. That gap between them is the entire congestion model in a single field pair.
What to watch for
The query engine is closed. Scry ships as a service rather than a repository, and the only public code is the agent skill pack under the ExoPriors org, MIT licensed. The corpus, the schema and the billing law all live on their servers, so you cannot point this at your own copy of the archive.
Nothing runs locally, which cuts both ways. The install really is two minutes and no code, but your questions travel to their infrastructure, and their security page asks for 90 days before publication on any vulnerability report.
Deadlines are short by default. A statement gets 15 seconds, raisable to as much as 2000 under an account key with x-scry-max-seconds, or 60 on the x402 lane. Long queries can start streaming a 200 response and then fail inside the body, so check for an error key before treating a response as rows.
The pricing story has drawn criticism. On the launch thread, commenters called the pricing pages hard to read and LLM generated, and asked the founder to explain it without invented terms. The founder’s answer was candid: queries are well under a cent per second when the server is not burdened, the price rises with load, and egress pricing is exponential because computation happens there rather than metaphorically.
Whether Reddit’s comment archive raises licensing trouble is also still open. The reply on HN was a two-word answer about pushshift, and nobody in that thread settled the question.
Scry is operated by Unflatten, a Berkeley company founded at the end of 2024. Asked on the launch thread why Scry was absent from the DeepSearchQA leaderboard for search agents, the founder said they had emailed the benchmark organisers, and that competing there needs no traditional search engine, only Scry or a direct fetch of a URL from a Scry result.
Source and Accuracy Notes
- Project page: https://scry.io/ (verified via direct visit, including the OG card)
- Documentation: https://scry.io/docs and https://scry.io/docs/connect (verified via direct visit)
- Live stats: https://api.scry.io/v1/stats/wire (53,956,370,500 rows, measured 2026-10-30)
- Price endpoint: https://api.scry.io/v1/scry/price (free_slack regime, congestion multiplier 0.0 at time of writing)
- Public agent skill: https://github.com/ExoPriors/skills (MIT license)
- Vulnerability disclosure: https://scry.io/legal/security (effective 2026-09-19)
- HN launch thread: https://news.ycombinator.com/item?id=49748041 (61 points, 2026-09-17)
Related Posts
dev-tools
Automotive Skills Suite for AI Engineering
Evaluate Automotive Skills Suite for APQP, ASPICE, HARA, safety-plan, and DIA workflows with setup notes, governance risks, and SME review guidance.
5/28/2026
dev-tools
awesome-agentic-ai-zh Roadmap Guide
Explore awesome-agentic-ai-zh as a Chinese agentic AI learning roadmap, with setup notes, track selection, study workflow, and evaluation guidance.
5/28/2026
dev-tools
Baguette iOS Simulator Automation Guide
Set up Baguette for iOS Simulator automation, web dashboards, device farms, gesture input, streaming, and camera testing with Xcode caveats.
5/28/2026