Introducing Scry: Programmable Internet Research via Congestion Pricing

Scry is currently in open alpha—a "one-person vibe coding" project designed to tame the vastness of the digital wilderness.

At its core, Scry is a Model Context Protocol (MCP) server. It empowers AI agents to execute complex programs across billions of web documents, transforming the internet from a collection of pages into a queryable database for both curiosity and commercial gain.

🌐 The Data Landscape

Scry indexes a massive array of high-signal sources:

• Community Hubs: Reddit, Hacker News, LessWrong, Stack Exchange.
• Knowledge Bases: Wikipedia, arXiv.
• Speculative Markets: Various prediction markets.

Millions of new entries are ingested daily. Instead of scraping a few pages, agents can pull thousands of rows or compute a final answer derived from every matching document in the corpus.

---

📊 Scale and Velocity

As of 19:15 UTC, the estate looks like this:

---

🚀 Performance Benchmarks

When tested against Google's DeepSearchQA (using 900 research questions with exact-set grading), Scry demonstrated superior precision:

The Result: An agent utilizing Scry achieved a 71.8% full-correctness rate. In comparison, Google's Kaggle leaderboard lists the Gemini Deep Research Agent at 66.1%.

Note: This was achieved using the same judge model and grading prompts. All transcripts are available on the DeepSearchQA page.

---

🛠️ How It Works

People search for pages $\rightarrow$ Agents query underlying records.

Instead of repeated searches and fragile scraping, Scry treats the web as a set of relations.

The Workflow:

1. Schema Discovery: The agent fetches the live schema via GET /v1/scry/schema.
2. Relation Selection: It identifies a source-native relation and any available vector helpers.
3. Bounded Execution: It writes a single query involving:
   • Phrase matching
   • Temporal windows
   • Joins
   • Vector composition

Each statement is read-only and constrained by a deadline, a memory ceiling, and a row limit.

Integration Checklist:

• Add MCP server: https://mcp.scry.io
• Compatible with ChatGPT
• Compatible with Claude / Claude Code
• Compatible with Codex / Cursor

---

🧬 Deep Dive: The Reddit Relation

The reddit.posts table acts as a near-census of submissions from 2005-06-23 to 2026-09-18.

Schema Definition:

The comment tree is stored in reddit.comments, joined via link_id = concat('t3_', id).

---

💻 Programmatic Examples

1. Graph Traversal (The "Who is Hiring" Problem)

To find the depth of a thread, Scry uses a fixpoint over the reply graph rather than a simple search.

The Query:

POST /v1/scry/query 
{ 
  "program" : { 
    "relations" : { 
      "seed" : { "bodies" : [[{ "ids" : [ "49522897" ]}]]}, 
      "thread" : { 
        "bodies" : [ 
          [{ "rel" : "seed" }], 
          [{ "rel" : "thread" }, { "edge" : "hackernews.children" }] 
        ]
      } 
    }, 
    "out" : [], 
    "depth" : 16 
  }
}

Result: 417 nodes reached in 354ms across 10 statements.

2. Vector Composition (Finding Concepts)

Can you find a failure mode without naming it? By averaging three descriptive vectors (metrics, classrooms, reward hacking) and ranking LessWrong posts while excluding the term "Goodhart's Law," Scry identified "specification gaming" and "fitness-seeking AIs" as top results.

POST /v1/scry/embed 
{ 
  "name" : "metric", 
  "text" : "Once a measure becomes a target, people improve the reported number instead of the outcome..." 
}

---

🎯 Mission & Access

Mission Statement: Scry is public-benefit infrastructure. It is subsidized for independent, non-commercial research and sustained by commercial partnerships. The goal is to enable search over public evidence where every result is fully traceable.

Access Tiers:

• Public: Standard queryable relations.
• Reviewed: Some sources are in the "held estate" but require approved access.
• Commercial: For custom builds, large-scale evals, or dataset access, contact hi@scry.io.

Feedback: Send requests to POST https://api.scry.io/v1/feedback (no credentials required).