Skip to main content
Glama

finos-mcp

Lets AI assistants look things up in three open financial-industry standards, accurately and without being able to change anything.

ci PyPI Python License tests coverage tools recall@5 hallucinated ids

Start with a question

What you want to do

Pick this server

See it work

Find AI governance risks and controls with traceable references

AIGF

Governance example

Understand a trade field or validate CDM JSON

CDM

CDM example

Find the right desktop intent for a financial instrument

FDC3

FDC3 example

Official MCP Registry (published v0.3.2): AIGF · CDM · FDC3. All three are local stdio packages, not public hosted endpoints.

Claude Code and Codex plugins: A community-maintained plugin can bundle all three servers and their guided AIGF, CDM, and FDC3 skills. See the plugin installation guide for marketplace setup and local runtime requirements; the existing manual MCP commands below still work.

Claude Code quick connect: With Claude Code and uv installed, run one of the following commands for the standard you need:

claude mcp add --scope user finos-aigf -- uvx --python 3.12 finos-mcp-aigf==0.3.2
claude mcp add --scope user finos-cdm -- uvx --python 3.12 finos-mcp-cdm==0.3.2
claude mcp add --scope user finos-fdc3 -- uvx --python 3.12 finos-mcp-fdc3==0.3.2

Run claude mcp list and then ask your assistant to use get_control (AIGF), describe_type (CDM), or suggest_intent (FDC3). The client setup guide also provides tested JSON configuration examples for Claude Desktop, VS Code and Copilot, with troubleshooting for Windows and macOS.

To see a real MCP call without a model or API key, clone this repository, run uv sync --all-packages, then try:

uv run --no-sync python examples/try_it.py aigf

Choose either of the other two servers in the tested example recipes. This is independent community software, not an official FINOS product.

Related MCP server: osdu-mcp

What is this?

Banks and other financial firms share a set of open standards published by FINOS, the Fintech Open Source Foundation. Three of them matter here:

  • the AI Governance Framework, a catalogue of the risks of using AI in finance and the controls that reduce them;

  • the Common Domain Model (CDM), a shared way to describe trades and what happens to them;

  • FDC3, the standard that lets desktop finance applications talk to each other.

An AI assistant asked about these standards will often answer from memory and get details wrong: an id that does not exist, a field that is not in the schema. This project gives the assistant the real thing to look up instead. It packages each standard as a small server that speaks the Model Context Protocol (MCP), the open protocol that Claude and other assistants use to call external tools.

With the servers connected, an assistant can answer questions like:

  • "Which controls reduce the risk of prompt injection, and what do they map to in the EU AI Act?"

  • "Is this trade JSON valid CDM? If not, where exactly is it wrong?"

  • "My app has a stock ticker selected. Which FDC3 actions can I offer the user?"

Each answer comes with an identifier or link the reader can check against the standard itself.

The three servers

Server

Standard

What an assistant can do with it

finos-mcp-aigf

AI Governance Framework

Look up 24 risks and 25 controls by their published ids, see which controls address which risks, cross-reference 13 outside frameworks (NIST, ISO 42001, EU AI Act, OWASP), and search in plain language

finos-mcp-cdm

Common Domain Model

Explore 767 CDM types and 1,143 JSON schemas, browse products and trade events, and validate trade documents with exact error paths

finos-mcp-fdc3

FDC3

Look up 19 standard actions ("intents") and 28 data types, check a data object, and suggest which actions fit it

All three are built on finos-mcp-core, a shared package that holds the safety rules described below. The full list of 29 tools is in docs/tools.md, and docs/demo.md shows a real session, including the CDM server finding three deliberate mistakes in a trade document.

Quick start

You need uv and nothing else. There is no account, API key or database to set up.

First-time setup: Follow the MCP client setup guide for copy-paste, version-pinned configurations for Claude Desktop, Claude Code, VS Code and GitHub Copilot. It also covers Windows and macOS setup, verifying a real tool call, and troubleshooting. The quick-start examples below pin the published, end-to-end-tested v0.3.2 packages.

Claude Desktop or Claude Code. Add this to your MCP configuration (full example) and restart:

{
  "mcpServers": {
    "finos-aigf": { "command": "uvx", "args": ["--python", "3.12", "finos-mcp-aigf==0.3.2"] }
  }
}

From the command line.

# talk over stdin/stdout, the way desktop assistants connect
uvx --python 3.12 finos-mcp-aigf==0.3.2

# or serve HTTP on localhost (self-hosted, not a public hosted service)
uvx --python 3.12 finos-mcp-aigf==0.3.2 --transport streamable-http --port 8000

As a container. It needs no writable disk and no extra privileges:

docker run --rm -p 8000:8000 --read-only --cap-drop ALL \
  -e FINOS_MCP_ALLOWED_HOSTS='localhost:*,127.0.0.1:*' \
  ghcr.io/vardhjain/finos-mcp:v0.3.2

The allowed-hosts setting makes the server refuse requests that arrive under any other host name, which stops a web page in your browser from reaching the port.

Running a stdio command directly starts a server waiting for a client; it does not display a chat prompt. To verify installation in a terminal, add --help instead. Swap finos-mcp-aigf for finos-mcp-cdm or finos-mcp-fdc3 to run the other servers. examples/ includes a low-memory, reproducible live demo, a short Python client, and a LangGraph agent.

Why it is safe to connect

Giving an assistant new tools is a risk, so the servers are deliberately limited:

  • They can only read. No tool writes a file, sends a message or changes any system. The server refuses to serve any tool if one of them is not marked read-only.

  • They work offline. The standards are copied into the packages at build time, with the exact upstream version recorded. Nothing is fetched while the server runs.

  • They cannot be flooded. Every tool has a rate limit and a cap on input and output size.

  • Mistakes are explained. A bad request gets a structured error that says what was wrong and what to try, so the assistant can correct itself instead of guessing.

  • Everything is logged. Each call is written to an audit log, with a fingerprint of its arguments in place of the raw content.

The details, including every setting, are in docs/safety.md.

One optional feature does use the network. Installing the semantic extra adds meaning-based search, which downloads a small open model on first use. Without it, search is keyword-based and fully offline.

How well it works

These numbers are measured by automated checks and published to the metrics branch. The badges above read that file directly, so they show measurements from a specific commit.

Check

Result

Automated tests

Over 200 passing, with more than 90% of the code exercised. The exact figures are on the badges above

Finding the right record for a plain-language question (50 questions)

Correct answer in the top five 91.5% of the time

An AI model answering 20 governance questions through the server

0 invented ids; 95% of answers cited a valid id

Response time per tool call

A few milliseconds on the benchmarked tools

The offline metrics update on every push to main; the search result in this table is from October 2026. The agent test uses Claude Haiku 4.5 and requires paid Anthropic API credits. To prevent surprise costs, it only runs through an explicit manual opt-in in GitHub Actions, never on a timer or PR label. Its published figures are from the last successful evaluation on 14 September 2026 and are not live data. Deterministic tests and retrieval evaluations continue to run in CI without LLM API calls.

Building this also turned up 15 defects in the upstream standards, recorded in docs/upstream-findings.md. The first was reported to FINOS, and its YAML fix merged upstream in PR #380 on October 21, 2026.

Project layout

Path

Contents

core/

Shared safety layer: read-only enforcement, rate limits, size caps, audit log, search

servers/

One package per standard, each with its copy of the upstream content

evals/

Golden tests, AIGF retrieval, offline CDM/FDC3 workflows, latency and optional agent evaluation

examples/

Configuration and sample clients

docs/

Documentation, also published at https://vardhjain.github.io/finos-mcp/

Development

uv sync --all-packages --all-extras
uv run pytest
uv run pytest evals/offline/test_workflows.py -v  # 16 multi-tool scenarios, no paid APIs
uv run ruff check . && uv run mypy core/src servers/*/src

See CONTRIBUTING.md for the ground rules, CHANGELOG.md for what changed in each release, RELEASING.md for how releases are made, and PLAN.md for the original design record.

Status

For maintainers, MCP directory and Registry status records the three already-published official Registry entries and the resolved Glama classification question. Future third-party directory submissions require separate approval.

The latest published release is available on PyPI as finos-mcp-core, finos-mcp-aigf, finos-mcp-cdm and finos-mcp-fdc3, for Python 3.12 and later, and as a container image for x86 and ARM.

Not affiliated with or endorsed by FINOS. These are independent community packages. Some official FINOS packages on PyPI also start with finos-; nothing here is one of them.

Licenses

The code is Apache-2.0. The copied standards keep their own licenses: the AI Governance Framework is CC-BY-4.0, and CDM and FDC3 are under the Community Specification License 1.0. See NOTICE.

Available Tools

11 tools
find_by_external_referenceA
Read-onlyIdempotent

Reverse crosswalk: find the AIGF risks and controls that cite an external reference, e.g. key='sa-9' (NIST SP 800-53), 'llm01-2025' (OWASP LLM Top 10), 'c3-s2-a15' (EU AI Act), 'A-6-2-6' (ISO 42001). Matching is case-insensitive and ignores punctuation; optionally restrict to one framework name from list_reference_frameworks.

Inputs: key (required), framework. Returns: key, matches, framework

get_controlA
Read-onlyIdempotent

Get one AIGF control by id (AIR-PREV-020, mi-20, 20, or its title), with the risks it mitigates, related controls, and crosswalk references (NIST SP 800-53, ISO 42001, ...). Set include_sections=false for a compact record.

Inputs: id (required), include_sections. Returns: AIGF calls these "mitigations"; we expose both words.

get_riskA
Read-onlyIdempotent

Get one AIGF risk by id (AIR-SEC-010, ri-10, 10, or its title), with the controls that mitigate it, related risks, and external references. Set include_sections=false for a compact record without the full text.

Inputs: id (required), include_sections. Returns: id, type, title, status, +10 more

list_controlsA
Read-onlyIdempotent

List AIGF controls (the framework calls them mitigations). Filter by type (PREV = preventative, DET = detective) or status. Returns ids for get_control.

Inputs: page, type, status, page_size. Returns: page, items, total, page_size

list_reference_frameworksA
Read-onlyIdempotent

List the external reference frameworks the AIGF crosswalks to, with entry counts and how many risks and controls cite each.

Inputs: no named parameters. Returns: frameworks

list_risksA
Read-onlyIdempotent

List AIGF risks. Filter by type (RC = regulatory and compliance, OP = operational, SEC = security) or document status. Returns ids you can pass to get_risk.

Inputs: page, type, status, page_size. Returns: page, items, total, page_size

map_control_to_externalA
Read-onlyIdempotent

Crosswalk one control to external frameworks: NIST SP 800-53 r5, NIST AI 600-1, ISO 42001, EU AI Act, OWASP LLM/ML/ASI, FFIEC, IOSCO, SR 11-7 and others. Optionally restrict to the named frameworks (see list_reference_frameworks).

Inputs: id (required), frameworks. Returns: refs, title, control_id, frameworks, +1 more

map_risks_to_controlsA
Read-onlyIdempotent

Map risks to the controls that mitigate them. Pass explicit risk_ids (up to 25), a free-text query (the top k matching risks are used), and/or up to 5 queries (each contributing its own top k matches, unioned with everything else) -- useful for "both X and Y" questions where a single query blurs together two distinct concepts. Controls are ordered by how many of the given risks they cover; control_type (PREV or DET) restricts the result. Reports uncovered risks and any inputs that did not resolve.

Inputs: k, query, queries, risk_ids, +1 more. Returns: edges, risks, controls, unresolved, +1 more

search_frameworkA
Read-onlyIdempotent

Search risks and controls with matching sections and citable source URIs.

For an unscoped question clearly asking for risks or for controls, rank those record types first while retaining the other type as a fallback. This is a stable partition of the original relevance-ranked results, not a new relevance score. Questions asking for both kinds stay mixed; explicit scope always wins.

Inputs: query (required), k, scope. Returns: hits, query, scope

search_statusA
Read-onlyIdempotent

Report search_framework's active retrieval mode: whether it is BM25-only (lexical) or fused with dense semantic search (hybrid), and why -- e.g. the semantic extra is not installed, or a model failed to load.

Inputs: no named parameters. Returns: Read-only introspection of `search_framework`'s retrieval mode.

server_infoA
Read-onlyIdempotent

Describe this server: standard version, upstream provenance, exposed counts, latency, policy.

Inputs: no named parameters. Returns: name, tools, counts, policy, +9 more

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.2/5.0

    Scored across 11 tools

    Disambiguation4/5

    Most tools target clearly distinct resources and actions (list_risks vs get_risk vs list_controls vs get_control). The main overlap is between search_framework and the list_* tools, and between map_control_to_external and find_by_external_reference, but the descriptions make the boundaries (forward vs reverse crosswalk, free-text search vs filtered listing) reasonably clear.

    Naming Consistency4/5

    Strong verb_noun pattern: list_risks, get_risk, list_controls, get_control, map_risks_to_controls, map_control_to_external, find_by_external_reference. Minor deviations are search_status and server_info, which describe state rather than an action but are still readable and predictable.

    Tool Count5/5

    11 tools is well within the ideal 3-15 range for a reference-framework server. Each tool earns its place: search, per-type listing and retrieval, bidirectional mapping, and reference listing/info all cover distinct needs.

    Completeness4/5

    Coverage is strong for a read-only reference surface: search, list/get for both risks and controls, risk-to-control mapping, control-to-external crosswalk, reverse crosswalk, and framework listing. The main gap is that external crosswalking is only directly exposed for controls (map_control_to_external), with risk-side external mapping only inferable via the reverse lookup.

    Maintenance

    ActivityNo data
    ResponsivenessOutdated data

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables AI agents to understand and query your database safely by providing a semantic layer of metadata, with tools to search, explain, validate, and generate safe SQL.
      2
      MIT
    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables AI agents to interact with the OSDU® data platform through Model Context Protocol, providing tools for search, storage, schema management, and dataset operations.
      2
      Apache 2.0
    • A
      license
      Not graded
      quality
      B
      maintenance
      Enables AI agents to query a repository's verified semantic model, including capabilities, ownership, contracts, constraints, impact, health, and verification status, without re-deriving project facts. It exposes read-only tools for inspecting, reasoning about, and verifying the provenance-tracked surface.
      2
      Apache 2.0