Overview

US bank data from the FDIC BankFind Suite API, which is public and needs no API key. Find any FDIC-insured bank or savings institution, including closed, merged, and failed ones, then read its quarterly Call Report financials back to 1984, rank it against peers, screen many banks at once, search failures since 1934, and map branch deposits and market share. Runs as a stdio process or a local Streamable HTTP server.

Institutions are keyed by their FDIC certificate number (CERT), which survives renames and charter conversions. Credit unions are insured by the NCUA and are not in this data.

Tools

Dataframe workflow

1. fdic_query_financials and fdic_get_deposits return an inline preview. When the full result is larger than limit, they also stage it as a df_XXXXX_XXXXX table and name it in the response's dataset field.
2. Pass that name to fdic_dataframe_describe for the table's columns and the unit and basis of each numeric column.
3. Query it with fdic_dataframe_query: joins, aggregates, window functions, and CTEs across every staged table. register_as saves a result as a new table.
4. Tables expire after FDIC_DATASET_TTL_SECONDS (default 24 hours), and the live tables hold 1,000,000 rows together: past that, the oldest are dropped to make room. With FDIC_DATAFRAME_DROP_ENABLED=true, fdic_dataframe_drop removes one sooner.

Staging runs on DuckDB and is on by default; CANVAS_PROVIDER_TYPE=none turns it off. The Claude Desktop .mcpb bundle ships without DuckDB's native binding, so there the dataframe tools report an install hint and every other tool works normally.

Capability reference

fdic_search_institutions _tool

• Filters: name (up to 100 characters; every word must match a current, former, or trade name), up to 50 certs, state, city, bank_classes, min_assets / max_assets (thousands of US dollars), holding_company_rssd
• status is active, inactive, or any; it defaults to any when name or certs is given and active otherwise, and the applied value comes back as status_filter
• Up to 100 per page (default 20), paged with offset / next_offset; records carry active, successor_cert for merged or failed banks, holding_company.rssd, and matched_on when a former or trade name matched; requested CERTs with no record land in missing_certs

fdic_get_institution_financials _tool

• cert required; up to 25 metrics (default: a 15-metric health set); quarters 1–200 (default 8), optionally bounded by from_date / to_date as 2025-03-31, 20250331, or 2025Q1
• Rows run most recent first, each a values map of metric → number, null when not reported; metric_definitions gives each metric's FDIC field, unit, and basis
• quarters_available counts the quarters in the window before the cap; an unknown CERT fails as cert_not_found

fdic_compare_peers _tool

• cert required; report_date defaults to the latest published quarter (report_date_defaulted); up to 20 metrics
• Peer group: peer_asset_band (same by default, any, under_100m, 100m_1b, 1b_10b, 10b_250b, over_250b) plus an optional peer_state, or an explicit peer_certs list of up to 200, which can't be combined with the band or state (conflicting_peer_filters)
• Each comparison returns the institution's value, peer_median, peer_p25 / peer_p75, peer_min / peer_max, percentile, and rank of rank_of, over the peer_count_with_value peers that reported the metric

fdic_query_financials _tool

• Filters: up to 100 certs, state, min_assets / max_assets, and up to 5 metric_filters (metric with min and/or max in the metric's unit); up to 30 metrics; from_date / to_date, defaulting to the latest published quarter only
• Inline preview of up to 500 rows (default 50), ordered by sort_by / sort_order; a panel larger than the preview is staged whole as a dataframe
• The panel is capped at FDIC_PANEL_MAX_ROWS (default 50,000) and keeps the newest quarters when it binds; total_matching, rows_fetched, and panel_truncated report what was left out

fdic_search_failures _tool

• Filters: name, up to 50 certs (events before 1977 carry none), state, from_date / to_date (YYYY-MM-DD), methods, min_assets; resolution is failure by default, or assistance or all
• summary totals every matching event, not just the page, with a by_method breakdown; group_by (year, state, method, insurance_fund) adds groups
• estimated_loss_missing_count counts events without an FDIC loss estimate, and a loss total is null when none has one; sort by date, loss, or assets, up to 200 per page (default 25)

fdic_get_deposits _tool

• cert alone returns the institution's branches and its deposit share in each state (mode: "institution"); a geography (state, county, city, zip, msa_code) ranks every institution in that market with hhi (market); both give the institution's position in the market (institution_in_market). county and city require state
• year from 1994 on, defaulting to the latest survey (year_defaulted); deposits are domestic branch deposits as of June 30, in thousands of US dollars
• Up to 200 branches or ranked institutions inline (default 25); a larger set is staged as a dataframe, and total_rows gives the full count

fdic_list_reference _tool

• One topic per call: metrics, bank_classes, failure_methods, insurance_funds, peer_asset_bands, or coverage
• metrics lists all 49 metric names with FDIC field, unit, basis, caveat, and in_default_set; served from built-in tables with no request to FDIC

fdic_dataframe_describe _tool

• name (a df_XXXXX_XXXXX from a dataset field) returns that table in full: source tool, the parameters it was called with, row count, expiry, column_schema, and column_units
• Without name, lists the live tables newest first, 50 per page, paged with offset / next_offset. Over HTTP with MCP_AUTH_MODE=none, where every caller shares one canvas, the listing is off and fails as listing_unavailable; a table is reached by its name
• Fails as canvas_unavailable when dataframe staging is off

fdic_dataframe_query _tool

• One SELECT per call, up to 20,000 characters; writes, DDL, file-reading functions, and system catalogs are rejected with typed reasons (non_select_statement, denied_function, system_catalog_access, and others)
• row_limit 1–10,000 (default 1,000) caps rows materialized, with row_count_capped when it binds; preview limits the rows returned inline
• register_as saves the result as a new dataframe with a fresh TTL, or fails as register_as_too_large when the result alone passes 1,000,000 rows; BIGINT results such as COUNT(*) come back as strings

fdic_dataframe_drop _tool

• name of a staged dataframe; returns dropped: true, or dropped: false when nothing by that name was staged
• Disabled unless FDIC_DATAFRAME_DROP_ENABLED=true; otherwise tables expire on their TTL

Features

Built on @cyanheads/mcp-ts-core: stdio and Streamable HTTP transports, pluggable auth (none / jwt / oauth), swappable storage (in-memory, filesystem, Supabase, Cloudflare KV/R2/D1), structured logging with optional OpenTelemetry tracing.

FDIC-specific:

• Reads four BankFind datasets: institutions (every insured charter, active and inactive), quarterly Call Report financials from 1984, failures and assistance transactions from 1934, and the annual Summary of Deposits from 1994
• Shared request pacer: request starts capped at FDIC_RATE_LIMIT_RPS, at most four in flight, a cooldown that holds every caller after an FDIC 429, and an in-process response cache; each call runs under one 45-second budget (55 seconds for a panel), so an error arrives inside a typical 60-second client timeout
• A curated catalog of 49 Call Report metrics with FDIC field code, unit, and basis; _ytd metrics accumulate from January 1, and unsuffixed income and return metrics cover a single quarter
• Handles BankFind's quirks: case-sensitive code values, string-typed report dates, absence sentinels, and ratios reported as 0 when a bank did not report them
• Peer statistics, deposit market share, and HHI are computed from every per-institution value, not from samples

Agent-friendly output:

• Provenance on every data response: data_as_of (the FDIC index build time), metric_definitions with unit and basis, and report_date_defaulted / year_defaulted flags when the server picked the period
• No fabricated values: unreported metrics come back null, and failure loss totals report how many events lack an estimate instead of counting them as zero
• Typed errors with recovery text that names the next call (cert_not_found, conflicting_peer_filters, no_report_for_period); rate-limit errors (pacer_shed, upstream_rate_limited) carry retryAfter
• Staged dataframes keep their provenance and per-column units, so SQL results stay interpretable

Getting started

Add the following to your MCP client configuration file. No API key is needed.

{
  "mcpServers": {
    "fdic-banks-mcp-server": {
      "type": "stdio",
      "command": "bunx",
      "args": ["@cyanheads/fdic-banks-mcp-server@latest"],
      "env": {
        "MCP_TRANSPORT_TYPE": "stdio",
        "MCP_LOG_LEVEL": "info"
      }
    }
  }
}

Or with npx (no Bun required):

{
  "mcpServers": {
    "fdic-banks-mcp-server": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@cyanheads/fdic-banks-mcp-server@latest"],
      "env": {
        "MCP_TRANSPORT_TYPE": "stdio",
        "MCP_LOG_LEVEL": "info"
      }
    }
  }
}

Or with Docker:

{
  "mcpServers": {
    "fdic-banks-mcp-server": {
      "type": "stdio",
      "command": "docker",
      "args": ["run", "-i", "--rm", "-e", "MCP_TRANSPORT_TYPE=stdio", "ghcr.io/cyanheads/fdic-banks-mcp-server:latest"]
    }
  }
}

For Streamable HTTP, set the transport and start the server:

MCP_TRANSPORT_TYPE=http MCP_HTTP_PORT=3010 bun run start:http
# Server listens at http://localhost:3010/mcp

Prerequisites

• Bun v1.4.0 or higher (or Node.js v24+).
• No account or API key: the FDIC BankFind Suite API is public.

Installation

1. Clone the repository:

git clone https://github.com/cyanheads/fdic-banks-mcp-server.git

2. Navigate into the directory:

cd fdic-banks-mcp-server

3. Install dependencies:

bun install

4. Configure environment:

cp .env.example .env
# every variable is optional; edit .env to override a default

Configuration

See .env.example for the full list of optional overrides.

Running the server

Local development

• Build and run the production version:

  sh

  # One-time build
  bun run rebuild
  
  # Run the built server
  bun run start:http
  # or
  bun run start:stdio
  

• Run checks and tests:

  sh

  bun run devcheck  # Lints, formats, type-checks, and more
  bun run test      # Runs the test suite
  

Project structure

Development guide

See CLAUDE.md for development guidelines and architectural rules. The short version:

• Handlers throw, framework catches — no try/catch in tool logic
• Use ctx.log for logging, ctx.state for storage
• Register new tools in buildToolDefinitions() in src/mcp-server/tools/definitions/index.ts
• Wrap external API calls: validate raw → normalize to domain type → return output schema; never fabricate missing fields

Contributing

Issues are welcome. Run checks and tests before submitting:

bun run devcheck
bun run test

License

This project is licensed under the Apache 2.0 License. See the LICENSE file for details.