Financial Data MCP Server for Quantitative Research

A financial data MCP server is most useful when it exposes dated research inputs, typed filters, and source fields rather than a generic stream of current company facts.

For quantitative research, a financial data MCP server should let an agent answer a narrow question with records that can be queried again. Alphanume's hosted server exposes 25 tools over point-in-time equity, options, event, regulatory, and reference datasets. The model can discover those tools and compose a workflow, while the data remains the same deterministic data returned by the REST API.

That distinction matters. MCP is a protocol for exposing capabilities to a model client, not a research result and not a market-data quality guarantee. The official MCP server specification separates tools from resources and prompts. Alphanume uses tools: each call sends filters to a read-only API route and receives structured JSON. The agent still needs instructions about dates, null values, and what claim it is allowed to make.

What the 25 tools cover

The Alphanume MCP documentation is the authoritative catalog and connection guide. Supported clients can use https://mcp.alphanume.com/mcp with browser OAuth. Clients without a browser flow use https://mcp.alphanume.com/key/mcp and send an API key through X-API-Key or a Bearer header. Both endpoints expose identical tools and enforce the caller's existing tier.

Run a point-in-time volatility screen

Consider the question: which stocks were optionable on a fixed date, had unusually rich settled implied volatility, and sat inside a known market regime? It spans three contracts. The optionable universe answers what could be traded then. IV/HV Premium answers how implied volatility compared with realized volatility. S&P 500 Risk Regime supplies a market-level label for the same session.

Use only Alphanume MCP tools.

For 2026-08-03:
1. Pull the historical optionable universe for that date range.
2. Pull settled IV/HV Premium rows with only_final=true and min_ratio_rank=0.9.
3. Pull the S&P 500 Risk Regime row for the same date.
4. Report truncated_to_max_rows for every data call.
5. If any call contains truncated_to_max_rows, stop and label the screen incomplete regardless of has_more. Write the equivalent REST handoff for a complete extraction instead of joining the partial rows.
6. Only if every input is complete, inner join optionability and volatility on date plus ticker.
7. Return date, ticker, has_weeklies, iv, hv, iv_hv_ratio,
   iv_hv_ratio_ranked, is_final, and risk_regime.
8. List unmatched rows separately and show every tool argument used.

This prompt does not ask the model to find a profitable options trade. A high ratio or high cross-sectional rank can reflect event risk, stale realized volatility, skew, liquidity, or a genuine premium. The screen creates a cohort for testing. It does not supply expected returns, transaction costs, or a position size.

Preserve a machine-readable audit record

A conversational answer is not enough for quantitative work. Save a compact record next to the returned table. The record should include the decision timestamp, tool names, exact arguments, access tier, truncation flags, and any join exclusions. This is the minimum needed to understand why a later rerun differs.

{
  "decision_date": "2026-08-03",
  "tools": [
    {"name": "get_optionable_tickers", "arguments": {"date_gte": "2026-08-03", "date_lte": "2026-08-03"}},
    {"name": "get_iv_hv_premium", "arguments": {"date": "2026-08-03", "min_ratio_rank": 0.9, "only_final": true}},
    {"name": "get_sp500_risk_regime", "arguments": {"date": "2026-08-03"}}
  ],
  "join_keys": ["date", "ticker"],
  "null_policy": "preserve",
  "truncation_checked": true
}

Large responses require care. Data tools return at most 500 rows through MCP, and that assistant-facing truncation happens after the REST response envelope is formed. When truncated_to_max_rows appears, inherited has_more and cursors cannot certify completeness. Stop the MCP screen and perform the wide optionable and volatility pulls through REST, following each route's REST cursors there. Do not treat the cap as a sample-selection rule.

Limits the server cannot remove

• An agent can join records from different dates unless the prompt forbids it.
• Current-state ticker classification has no date dimension and should not be described as historically versioned.
• A free account's trailing 20-session delayed window is suitable for testing mechanics, not a long-horizon claim.
• MCP OAuth simplifies credential handling, but authorization does not certify a model's interpretation.
• Point-in-time inputs reduce leakage only when forward outcomes are attached after the cohort is frozen.

Start by connecting through the documented endpoint, call check_api_status, and run the fixed-date prompt above. Save the audit JSON and the result table before changing any filter. Compare adjacent workflows in the agent-research topic hub and the systems argument in A Model for Quant Research at Industrial Scale. If the workflow needs deeper history, confirm access on the pricing and access page; if it needs automation, translate the accepted calls into REST and test response parity before scheduling them.