DIRECT
Describe a setup in the browser, inspect the exact translated filters, run the scan, and open any occurrence in replay.
Market microstructure search, built for agents
Search 33 microstructure features across 660+ Binance USDT-M crypto and TradFi perpetuals. Use the same deterministic engine directly, through the REST API, or as grounded tools your own AI agent can call. Every count includes its denominator, every outcome is computed forward, and every result carries replayable evidence.
Read-only v1. Research describes the recorded past, not the future. Nothing here is trading advice.
ONE ENGINE · THREE WAYS IN
Search it yourself, call it from code, or delegate it
REST API
Submit the same versioned query documents from a notebook, application, or automated research workflow.
MCP FOR AI AGENTS
Give Claude, Cursor, Codex, or any MCP client grounded tools for registry discovery, interpretation, occurrence scans, pagination, snapshots, prevalence, commonality, cohorts, and reports.
ANSWERS WITH RECEIPTS
Three promises your agent can rely on
COUNTS WITH DENOMINATORS
Every count arrives with what it is out of: eligible buckets and warmup exclusions ride inside the result. A rate is never a bare number.
OUTCOMES NEVER FILTERABLE
Outcome fields are computed forward from each anchor and can never be a filter. You get the true rate over every occurrence, not a hindsight-selected list.
SAME KEY, SAME BYTES
Every result carries a reproducibility key. Quote it and anyone reproduces byte-identical bytes over the same record. Cache hits and ETag revalidations do not spend another scan unit.
MCP · ONE LINE
Point your agent at it
The MCP server exposes the same research primitives as agent tools, so Claude, Cursor, Codex and any MCP client can search the record and get answers with receipts. The agent grounds itself with list_features, then submits only fields and operators from the closed registry. The server holds no secrets: it forwards your key and streams the engine's canonical bytes back untouched.
npx -y @edgedepth/research-mcpTwo ways in: a hosted Streamable HTTP endpoint at https://mcp.edgedepth.com/mcp with no install, or the npx shim above running the same tools over stdio. One package, two transports.
Fixed-anchor stratification is the next tool in rollout, not part of the current public API or MCP tool set. It will move through direct-backend, byte-preserving public-route and MCP release gates before it is advertised as live.
ASK IT THINGS LIKE
Search across crypto and TradFi perps
A good question names a setup and lets the tool report what followed over every occurrence:
- When were top traders at least 15 percentage points more long than the crowd, and what happened over the next one and four hours?
- Find every market where at least $5M was liquidated in an hour while open-interest velocity was unusually high.
- Compare this setup across NVDAUSDT, QQQUSDT, XAUUSDT and XAGUSDT, then give me the exact moments to replay.
One thing it refuses by design: you cannot filter to the occurrences that then went up. Ask for that and you get OUTCOME_IN_PREDICATE back, so the rate you see is always the real one, over every occurrence.
Give your agent a market record it can cite.
33 FEATURES · DETERMINISTIC · REPLAYABLE · READ-ONLY V1
Research describes observed occurrences in the selected record window, not a guarantee. Read-only v1. Nothing here is trading advice.