Research from the browser, API and MCP

EdgeDepth Research is not three separate products. The browser, public API and MCP server are clients of the same engine. They use the same closed 33-feature registry, the same research_query.v2 documents, the same frozen validator and the same result encodings. A query does not change meaning when it moves from a person clicking RUN to an application making an HTTPS request or an agent calling a tool.

Run it in the browser

The Research workspace is the most visual way in. Type a condition in plain words or point at a market minute. The interpreter proposes exact chips, and every inferred value blocks execution until you confirm or edit it. Nothing runs on prose alone.

The browser also carries the stateful product workflow around a result:

• inspect occurrences and forward outcome distributions;
• compare a where-only matched cohort with the eligible baseline;
• read a full 33-feature snapshot as of one minute;
• compare up to eight selected moments for deterministic commonality;
• publish a citable report;
• arm a verified result as a standing live alert;
• open any occurrence or alert fire in tick-by-tick replay.

Use the browser when a human needs to inspect, edit, compare or arm the result.

Call the public API

The Research API exposes the same work over HTTPS with API-key authentication. Start by reading the registry, then submit versioned documents rather than inventing feature names. The main surfaces cover registry discovery, interpretation without execution, scans, cohort comparisons, snapshots, commonality and published reports.

The API preserves the engine contract. Counts carry denominators, validation failures return stable machine-actionable codes, absent values remain absent, and canonical results include the reproducibility metadata needed to rerun or cache them. Prose interpretation and execution are deliberately separate calls: receiving a proposed document is not permission to run it.

Use the API when your own software already knows the question and needs a stable JSON/HTTP boundary. See the grammar reference and worked examples.

Connect an AI agent with MCP

The EdgeDepth MCP server wraps that public API as nine read-only tools:

• list_features
• interpret_prose
• run_scan
• run_cohort
• next_page
• snapshot_at
• base_rate
• commonality
• get_report

Connect a client that supports bearer headers to the remote Streamable HTTP server at https://mcp.edgedepth.com/mcp, or run the published @edgedepth/research-mcp package locally over stdio:

npx -y @edgedepth/research-mcp

The local process reads EDGEDEPTH_API_KEY; it does not store your key or rewrite result bytes. The MCP setup guide has configuration examples for Codex, Cursor, Claude and local clients.

The agent must still follow the confirmation boundary. interpret_prose returns a proposed document. The agent should present its exact fields, thresholds and time range to you before calling run_scan or run_cohort. The MCP tools are read-only: they can research and retrieve, but cannot arm alerts, alter account settings or place a trade.

Which surface should you use?

Use the browser to explore and make judgments. Use the API to integrate the engine into software. Use MCP when you want an agent to ground itself in the registry, build an explicit query and retrieve the canonical evidence on your behalf. You can move the same document between all three without translating its meaning.

The important object is not the client. It is the receipt: query grammar, feature version, dataset revision and canonical query hash. Same receipt, same result bytes.

---

Occurrences describe only the recorded markets, dates and eligible buckets reported by the scan. They are observed examples, not a guarantee. Nothing here is trading advice.

> Open Research