EdgeDepth Research

EdgeDepth Research MCP Server

@edgedepth/research-mcp is the official, research-only Model Context Protocol server for EdgeDepth, a market microstructure search engine over recorded Binance USDT-M crypto and TradFi perpetuals. Use it from ChatGPT, Claude, Cursor, Codex, or any MCP client to find every verified occurrence of a market condition, inspect forward outcomes across the complete matched set, read an unconditional same-scope reference, and open replay-linked evidence.

Every result includes counts with denominators and a reproducibility key. Same key, same bytes.

Website · Search the market · REST API documentation · MCP setup guide · Learning hub

Why use EdgeDepth Research?

• Search recorded market microstructure: query a closed, versioned feature registry covering order flow, price action, volatility, funding, open interest, positioning, candle formations, and liquidations.
• Keep the denominator: every count reports the eligible population and exclusions behind it. Missing data is absent, never silently changed to zero.
• Measure outcomes without lookahead selection: forward returns, MFE, and MAE are computed over all occurrences. Outcome fields cannot be used as filters.
• Compare matched and baseline populations: deterministic cohort results put the matched distribution beside every other eligible predicate-false bucket.
• Audit and replay the evidence: results carry a reproducibility key, and representative occurrences include authenticated web handoffs to the exact recorded market moment.
• Stay research-only: no tool trades, modifies alerts, publishes reports, or writes account data. A fresh scan, cohort, or stratified computation can consume research allowance units; the annotations state that side effect explicitly.

Choose a connection

The package exposes one tool core through two transports:

• Hosted MCP (recommended): connect to https://mcp.edgedepth.com/mcp over Streamable HTTP and authorize once in your browser. No API key to copy.
• Local stdio: run npx -y @edgedepth/research-mcp with an EdgeDepth API key.

Connect

Claude Desktop

In Settings > Connectors > Add custom connector, enter:

https://mcp.edgedepth.com/mcp

Complete the EdgeDepth browser authorization prompt.

Cursor (~/.cursor/mcp.json)

{
  "mcpServers": {
    "edgedepth-research": {
      "url": "https://mcp.edgedepth.com/mcp"
    }
  }
}

Codex (~/.codex/config.toml)

[mcp_servers.edgedepth]
url = "https://mcp.edgedepth.com/mcp"

Then run:

codex mcp login edgedepth

Remove any old bearer_token_env_var line before using browser OAuth.

Local stdio with npx

Create a key on the EdgeDepth Developer page, then add:

{
  "mcpServers": {
    "edgedepth-research": {
      "command": "npx",
      "args": ["-y", "@edgedepth/research-mcp"],
      "env": {
        "EDGEDEPTH_API_KEY": "edk_live_YOUR_KEY"
      }
    }
  }
}

Local stdio requires Node.js 20 or newer. Use the research:read key scope for recorded-data tools and add research:interpret only when you need the free interpret_prose proposal step.

Result projection (agent context economy)

Scan-family results are large: a universe scan's canonical bytes run to hundreds of kilobytes, most of it page rows carrying every recorded feature, the zero and long-tail entries of counts_by_symbol, and empty threshold rungs. That overflows a client's tool-result budget before it answers anything.

run_scan, next_page and run_cohort therefore return a stated projection by default. It only ever REMOVES, and every removal is listed in a trailing note with the exact way to get the bytes back:

• occurrence rows are trimmed to rows (default 3) and each kept row keeps the setup fields its own evidence block names - full_rows: true restores the whole vector;
• the per-occurrence outcomes map keeps the entries for the rows that remain;
• counts_by_symbol keeps the top entries by match count, and says how many instruments and matches were omitted;
• the outcome ladders are replaced by a paired answer block: for each metric, present, absent and the selected rungs' integer counts pass through verbatim, with rate, the unconditional baseline_rate over the same symbols and window, and their ratio as lift stated beside them. The selection is fixed in advance (gte 0.01, gte 0.02, lte -0.01, lte -0.02), drops rungs that separate nothing, and adds the single rung carrying the largest lift among those holding at least 30 occurrences, marked kept_for. full_outcomes: true returns every rung and the per-rung histogram, on the matched set and the reference separately.

Counts, denominators, absent tallies, predicate_coverage, representatives, the page cursor and the reproducibility key are never touched, and the request document is never rewritten, so the canonical query hash and the credit charged are exactly what you asked for. full_counts: true returns the engine's verbatim canonical bytes with no projection at all. ETags are projection-scoped: an ETag held for one projection can never revalidate as a different one.

list_features takes the same treatment on request: search, feature_ids and compact return one feature family instead of the whole grammar, with the closed parts (operators, windows, sequence rules, limits, error codes) intact.

Prompts and resources

The server publishes worked prompts, which compatible clients surface as pickable commands: test_a_claim, liquidation_cascade_bounce, investigate_symbol, does_it_confirm and how_common_is_it (the free prevalence path). Each one encodes the same answer contract: ground the grammar, propose the exact definition, wait for confirmation, then report with denominators, the reference, the reproducibility key and a replay handoff.

The grammar registry is also served as a resource, edgedepth://research/grammar, so a client can attach it once instead of calling list_features every session.

Recommended agent workflow

1. For natural-language questions, call interpret_prose first with the user's question unchanged. The interpreter already uses the registry. Do not insert unstated thresholds, dates, markets or outcome definitions. Use list_features only when constructing or repairing a document. It is the live, closed grammar and prevents invented fields. Its result also carries the human reading page for any feature id: https://edgedepth.com/research/readings/<id without the "feature." prefix>, so feature.vpin is explained at edgedepth.com/research/readings/vpin. Open it when a person needs to know what a reading measures before a threshold is chosen.
2. Call list_instruments only when you need to check the manifest-derived universe, coverage, and provenance. Its result carries the human market page in the same way, https://edgedepth.com/research/symbols/<symbol>, for a market still being recorded; a delisted market in the universe has no page, so offer that link rather than promising it.
3. Show one short proposal: condition, exact markets and dates/time zone, outcome definition and horizon, and metering. Interpretation is free; fresh computations can consume allowance. Label every unprovided value as a proposed assumption using chip provenance. Resolve unsupported fragments and ask only questions that materially change the study. Keep exact JSON and diagnostics inspectable in tool details, available on request.
4. Wait for explicit human approval, then pass the same document to run_scan. Changes require a new proposal and confirmation. The exact-document API does not store a proposal ID or a human approval receipt; client consent is required, and a model-supplied flag is not proof. Do not link an unapproved proposal through rq: that workbench handoff can execute on arrival.
5. Answer the question first, preserving zero-match and inconclusive findings. Give matched/eligible counts, coverage exclusions, present/absent outcomes, both directions at the agreed horizon, and overlap/selection limitations. Read rates from outcomes_summary, which covers all occurrences. Page rows are examples, never the denominator. Each rung already carries its matched count and rate, the unconditional rate, and their ratio as lift: quote those, and quote the count beside the rate. No lift means no reference was available or the unconditional rate was zero; neither licenses estimating one.
6. Read the appended unconditional same-scope reference when available. It is not matched, comparable, or a causal control.
7. Return the full reproducibility key with the answer and one relevant next action: a returned replay, a changed assumption, or an existing report. Saving and alerts remain web actions. Each handoff states how far back it sits; replay reach is a per-account entitlement, so an old moment can be refused at the web surface even though the occurrence is real. Use next_page only with a cursor returned by the API.

Example instruction for an MCP client:

Did elevated VPIN and one-sided buying tend to precede a rise? Propose a precise
study before running anything. Label any suggested thresholds, markets, dates
and outcome definition so I can approve or change them.

The user does not need tool names, feature IDs or JSON. The client translates the confirmed proposal into the existing exact-document call.

Tools

No tool can trade, change market state, publish, or modify account data. run_scan, run_cohort, run_stratified and outcome_first are annotated as metered computations because a fresh call can irreversibly consume an allowance unit. The other recorded-data tools are closed-world reads. interpret_prose is a free read that uses the configured external language interpreter.

Research contract

• Validation failures pass through as 422 {"errors":[{"code":"...","message":"..."}]}.
• Transport failures use the {"error","code"} envelope.
• Contract codes are machine-actionable. For errors such as UNSUPPORTED_FEATURE or OUTCOME_IN_PREDICATE, call list_features, repair the document, and retry.
• Deterministic tools are exact-document, UTC-only tools. interpret_prose may use a time zone to plan dates, but run_scan, run_cohort, and base_rate never reinterpret calendar language.
• Reruns and ETag 304 Not Modified revalidations are free. list_instruments ETags are scoped to the requested summary, symbol projection, or full representation.
• Interpretation is free and never debits the scan allowance. An unavailable scan allowance returns neutral 402 RESEARCH_ALLOWANCE_EXHAUSTED metadata without a checkout link.

REST API and documentation

The MCP server is a thin, deterministic interface to the public EdgeDepth Research API:

• REST API quickstart
• Authentication and API keys
• Credits, caching, and limits
• Versioned query grammar
• Reproducibility contract
• Worked API examples
• MCP connection guide
• How EdgeDepth Research works
• What you can ask

The default REST base used by the stdio package is https://app.edgedepth.com/api/v1/research.

Environment

Local stdio

Hosted server operators

Authentication and security

The hosted server uses browser OAuth. It validates opaque access tokens, exchanges them for separate short-lived internal assertions, and never passes the OAuth access token to the REST API. The MCP server is stateless and stores no user credentials.

Compatible clients rotate refresh tokens silently while the connection remains active. Review or revoke access at EdgeDepth Connected Apps.

API keys remain available for scripts, local stdio, and MCP clients without browser OAuth. Treat an edk_live_... key as a secret and never commit it to source control.

Develop

npm install
npm run build
npm test
npm run typecheck

TypeScript builds to dist/. Example nginx locations, systemd hardening, and operator environment values live under deploy/. Production deployment and npm publishing remain operator actions.

Related projects

• edgedepth-terminal (AGPL): the open-source C++/WASM orderflow terminal. Replay-linked evidence from research results opens the exact recorded market moment in it, and it self-hosts with one docker compose command.
• edgedepth-gateway (MIT): a Go bridge from Binance's public streams to the terminal's wire format, for running the terminal on live data without an account.

License

MIT

Inline scan evidence

Supported MCP Apps hosts can display a comparison and recorded-distribution card from run_scan. The card receives only complete-result forward-return summaries, coverage, exact query/key and metering in tool-result _meta. This data is hidden from the model in ChatGPT; the existing text projection is unchanged. No raw page observations are used to make distributions, no fitted curves are invented, and no additional requests or allowance consumption occur when changing chart views. Reference distributions are compared only when their bin edges align. Empty bins, open tails, missing outcomes and zero/one-observation states remain visible. Horizon and move-size controls are display choices over already-computed outcomes, not changes to the approved query. The card defaults to the labelled 1h / 1% view. Exact study/evidence details expand inside the card; text-only hosts keep the existing response. The HTML resource has no network dependencies or mutations. This is a developer-connector update, not an automatic official V1 rescan.