Scry Launch: Programmable Internet Search with Congestion Pricing

Scry enables agents to run arbitrary, bounded SQL‑style queries over a live, multi‑source web corpus

Takeaway: Scry provides a single MCP endpoint (https://mcp.scry.io) that lets ChatGPT, Claude, Codex, Cursor and any MCP‑compatible client execute read‑only queries across 43 public data sources covering 416 billion rows, with pricing based on the declared execution time of each query.

What Scry actually offers

Programmatic search: Agents can query Reddit, Hacker News, LessWrong, arXiv, Stack Exchange, Wikipedia, prediction markets, and many other sources using a SQL dialect that supports phrase matching, time windows, joins, and vector composition.
Live schema discovery: GET /v1/scry/schema returns a machine‑readable contract describing every relation, its columns, row counts, and freshness lag. Agents must select relations from this catalog; unavailable names are omitted.
Row‑level provenance: Results retain source‑native identifiers and timestamps, enabling traceability back to the original document.
Bounded execution: Each query runs under a deadline, memory ceiling, and row‑cap. Agents can request an execution plan via the x-scry-explain: 1 header to see estimated rows, bytes, and compute before paying.
Vector primitives: Embeddings are first‑class values. Functions such as scry_centroid, scry_contrast_axis_balanced, and scry_cosine_similarity let agents perform arithmetic on vectors and rank corpora along custom semantic axes.
Reranking: The x-scry-rerank header (or the /v1/scry/rerank endpoint) lets agents reorder already‑fetched rows with a local LLM at zero cost.

Scale and freshness as of 2026‑09‑11

Queryable rows: ~161 billion rows are immediately queryable across 43 sources.
Total held estate: 416 billion rows, with a daily ingest rate of +32.6 billion rows (≈22.7 M rows/min).
Key source breakdown (selected examples):
- Reddit comments: 27 billion rows (near‑complete archive from 2005‑present).
- Hacker News items: 45 million rows, refreshed within 15 minutes.
- Common Crawl pages: 20.8 billion rows of extracted text.
- OpenAlex works: 511 million rows of scholarly metadata.
- Code & dependency graphs: 43.9 billion rows across GitHub events, deps.dev, and package registries.
Undisclosed sources: 255 billion rows are counted in the held estate but are not queryable without special access.

Pricing model – “congestion pricing”

Tier	Cost	Intended audience
Researcher	$0 (includes $5 signup credit)	Non‑commercial, individual researchers
Patron	$100 / month (rolls over)	Hobbyists and small teams
Team	from $2,000 / month	Commercial use, dedicated capacity
Agents	$0.05 / second declared	Pay‑per‑use for autonomous agents

Agents are billed only for the time they declare in the query header (x-scry-max-seconds). This model discourages overly broad scans that would consume large fractions of a source’s row‑budget, effectively pricing congestion rather than raw data volume.

How to connect

ChatGPT – Enable Developer mode, add a plugin named scry pointing to https://mcp.scry.io, and sign in.
Claude – Add a custom connector with the same URL under Settings → Connectors.
Developers – Use the HTTP API (https://api.scry.io/v1/scry/query) with an API key from the dashboard, or any MCP client (Claude Code, Codex, Cursor).

Typical request (curl example):

curl -s https://api.scry.io/v1/scry/query \
  -H "Authorization: Bearer $SCRY_API_KEY" \
  -H "Content-Type: text/plain" \
  --data "SELECT hn_id, title FROM hackernews.items WHERE hn_id >= (SELECT max(hn_id) FROM hackernews.story_scores WHERE observed_on >= today() - 7) - 100000 ORDER BY hn_id DESC LIMIT 20"

Example queries that showcase Scry’s power

Find small‑followed accounts that bookmark more than they like (4.9 s over 204 M tweet revisions).\
Measure alignment drift on LessWrong using a contrast axis between two stance sentences (≈0.1 s).\
Identify common followees of Elon Musk, Sam Altman, and Eliezer Yudkowsky (171 ms over 2.9 M rows).\
Locate papers that cite both Scaling Laws and Chinchilla (0.1 s, 195 matches).\
Detect recurring feuds on Hacker News (1.6 s over 46.7 M rows).

These examples demonstrate how a single declarative statement can replace dozens of manual scraping and post‑processing steps.

Community feedback from Hacker News

codexon: “How did you scrape Reddit comments? Doesn’t this require expensive licensing?” – The post does not disclose licensing details; the comment highlights a common concern about data provenance.

ashkankiani: “Pricing reminds me of algorithmic trading. Could you distribute datasets via P2P like Wikipedia?” – Suggests a desire for open‑source distribution to reduce query‑side congestion.

vova_hn2: “Pricing model is hard to understand; ‘second of query time’ is undefined.” – Confirms that the novel pricing terminology may need clearer documentation.

DylanMerigaud: “Congestion pricing for queries sounds innovative.” – Positive reception of the economic model.

MrDrMcCoy: “Could this supplement smaller search engines and break Google/Bing dominance?” – Highlights potential ecosystem impact.

Overall sentiment is enthusiastic but calls for clearer explanations of data licensing, pricing mechanics, and potential open‑distribution of the underlying corpora.

Why Scry matters

Enables true programmatic research: Agents no longer need to stitch together page‑by‑page scrapes; they can ask the web a single, composable query and receive structured, provenance‑rich results.
Reduces redundant crawling: By exposing the raw source tables, Scry eliminates the need for each downstream tool to maintain its own crawler, lowering bandwidth and storage costs.
Introduces economic incentives: Congestion pricing discourages wasteful, high‑volume scans, aligning resource usage with the value of the insight gained.
Public‑benefit infrastructure: Free tiers for non‑commercial researchers lower the barrier to large‑scale data analysis, while commercial engagements fund the service.

Next steps for users

Sign up for a free researcher account to obtain an API key.
Retrieve the live schema (/v1/scry/schema) and explore available relations.
Prototype a query using the x-scry-explain header to gauge cost before execution.
Integrate the MCP endpoint into your LLM‑powered agent or analytics pipeline.

All performance numbers (e.g., 0.86 s per GB of Hacker News text) are measured on the live production system on 2026‑09‑11 and reflect single‑server execution under typical load.