camt053-mcp: An MCP Server for ISO 20022 Bank Statements

A Model Context Protocol server that exposes the camt053 ISO 20022 Bank Statement library as tools for AI agents and assistants — discover message types and return reasons, inspect input schemas, validate records and financial identifiers, parse incoming statements, and generate validated reversing-entry XML, all from your favourite MCP client.

Latest release: v0.0.14 — OAuth 2.1 resource-server auth (RFC 9728) on the HTTP transport, Prometheus metrics, a tamper-evident audit chain, and real-HTTP load benchmarks; 22 MCP tools over stdio or authenticated streamable HTTP, all backed by the shared camt053.services layer, for Python 3.10+. See what's new →

Contents

• Overview
• The ISO 20022 MCP Suite
• Install
• Quick Start
• Tools
• Prompts
• Resources
• Rulebook search
• Using the tools
• The camt053 suite
• When not to use camt053-mcp
• Development
• Security
• Documentation
• License
• Contributing
• Acknowledgements

Overview

The Model Context Protocol (MCP) is an open standard that lets AI agents and assistants discover and call external tools in a uniform way. camt053-mcp is an MCP server that turns the camt053 library into a set of first-class agent tools, so an assistant can read and reverse ISO 20022 camt.05x cash-management messages — the standardised bank-to-customer account reports, statements, and debit/credit notifications — directly from a conversation.

The headline capability is the one-shot reversing-entry workflow: read an incoming camt.053 statement, find the entries carrying a return reason code (e.g. AC04 Closed Account), and emit a validated reversing entry.

Every tool is a thin, typed wrapper over camt053.services — the single shared facade also used by the CLI and REST API — so all interfaces behave identically. Tools return JSON-serialisable data; on an error they return an {"error": ...} payload rather than raising.

• Website: https://sebastienrousseau.github.io/camt053/
• Source code: https://github.com/sebastienrousseau/camt053-mcp
• Bug reports: https://github.com/sebastienrousseau/camt053-mcp/issues

This package is part of the camt053 suite — a set of independently installable packages that share the camt053.services layer:

• camt053 — the core library (CLI + REST API)
• camt053-mcp — this package, the Model Context Protocol server
• camt053-lsp — the Language Server Protocol server for editors

flowchart LR
    A["MCP client<br/>(Claude Desktop, IDE, agent)"] -->|stdio| B["camt053-mcp"]
    B -->|delegates to| C["camt053.services"]
    C -->|parse + reverse + validate| D["ISO 20022 camt.053 XML"]

The ISO 20022 MCP Suite

camt053-mcp is the bank-statement flagship of eight coordinated, vendor-neutral MCP servers that together cover the ISO 20022 bank-statement workflow and the November 2026 structured-address cutover — statement depth, whole-catalogue routing, reconciliation, multi-format ingestion, and address remediation. Dependency ranges are kept aligned across the suite, so the servers co-install cleanly in a single Python environment: start with one, add the rest as your workflow grows.

In one line each: camt053-mcp is the bank-statement flagship (deepest camt.05x surface, stdio + authenticated streamable HTTP); iso20022-mcp is the generic message toolkit (a handful of verbs over the whole catalogue); reconcile-mcp is the reconciliation workflow (did the money we expected actually arrive?); bankstatementparser-mcp is the ingestion layer (many formats in, one transaction shape out); and structured-address-fix-mcp is the postal-address specialist (debtor/creditor addresses cliff-ready for the Nov 2026 cutover).

The suite also includes per-family servers — pain001-mcp (credit transfer initiation), pacs008-mcp (FI-to-FI credit transfers), and acmt001-mcp (account management) — reachable through the iso20022-mcp gateway.

Install

camt053-mcp runs on macOS, Linux, and Windows and requires Python 3.10+ and pip. It pulls in the core camt053 library and the MCP SDK automatically.

python -m pip install camt053-mcp

python -m venv venv
source venv/bin/activate        # macOS/Linux
venv\Scripts\activate           # Windows
python -m pip install -U camt053-mcp

Quick Start

For the 10-minute install → MCP client config → first conversation tutorial, see docs/quickstart.md.

Launch the server over stdio (the FastMCP default transport):

camt053-mcp

Register it with any MCP client (e.g. Claude Desktop) by adding it to the client's configuration:

{
  "mcpServers": {
    "camt053": { "command": "camt053-mcp" }
  }
}

The agent can then call the tools below to parse incoming statements and generate validated reversing entries on demand.

For a shared, multi-tenant deployment, the server can also serve streamable HTTP with mandatory bearer-token auth and optional per-request Camt053-Account tenant scoping:

CAMT053_MCP_TOKEN=<secret> camt053-mcp --transport=http --bind=0.0.0.0:8080

See Multi-tenant HTTP deployment and the deployment cookbook.

Tools

All tools delegate to the shared camt053.services layer, so they behave identically to the CLI and REST API.

• list_message_types — List the 3 supported camt.05x message types
• list_return_reasons — List the ISO external return reason codes
• get_required_fields — Required input fields for a message type
• get_input_schema — Full input JSON Schema for a message type
• validate_records — Validate flat records against a message type
• validate_identifier — Validate an IBAN, BIC, or LEI
• validate_statement — Validate a statement against its XSD and detect its type
• convert_mt940_to_camt053 — MT940 → camt.053 migration: convert legacy SWIFT MT940 statement text into a camt.053 structure
• convert_mt942 — MT942 → camt.052 migration: convert legacy SWIFT MT942 interim transaction report text into a camt.052 structure
• check_cbpr_readiness — Flag CBPR+ Nov 2026 cliff issues in a statement
• get_cbpr_cutover_date — Return the official CBPR+ cutover date (2026-11-16)
• cite_rulebook — Quote a curated SEPA / CBPR+ / HVPS+ rulebook clause
• list_rulebook_clauses — List the available rulebook citations (optionally filtered)
• search_rulebook_vector — Find rulebook clauses by natural-language similarity when you do not know the clause id (needs the [vector] extra)
• export_journal — Export statement entries as Xero BankTransactions or QBO JournalEntry payloads
• list_export_journal_targets — List the accounting-platform targets export_journal supports
• classify_entry — Classify a statement entry via MCP Sampling (uses the client's LLM)
• list_classify_entry_categories — List the default categories classify_entry uses
• get_tenant_context — Report the multi-tenant scope of the call (the Camt053-Account header on the HTTP transport; None over stdio)
• parse_statement — Parse an incoming camt.05x statement into data
• detect_statement_anomalies — Screen a statement for duplicate references, unusual fee deductions, and velocity spikes (deterministic rules, no model or network)
• list_entries — List every entry across all statements (paginated)
• filter_entries — Return entries carrying a return reason code (paginated)
• generate_reversal — Generate a validated reversing-entry XML document

Pagination

list_entries and filter_entries accept optional offset (default 0) and limit (default None) parameters. When limit is omitted they return the full list, exactly as before. When limit is given they return a paginated envelope instead:

{"total": 42, "offset": 10, "limit": 5, "entries": [/* ... */]}

A negative offset or limit returns an {"error": ...} payload, consistent with the rest of the server's error convention.

Prompts

reversal_preview takes an optional reason_code (default "AC04") and returns a four-step message template: parse the statement, preview the matching entries with filter_entries, confirm with the operator, then call generate_reversal. The other three prompts take no parameters and return a two-message user-prompt + assistant-walkthrough template the agent can replay verbatim.

Resources

Resources give an agent read-only reference context it can load without calling a tool. Each resource returns a JSON payload.

Both back onto the shared camt053.services layer, so they stay in sync with the equivalent list_return_reasons / list_message_types tools. On an error they return a serialised {"error": ...} payload.

Rulebook search

cite_rulebook needs a scheme, version and clause id. When you know what a rule is about but not what it is called, search_rulebook_vector closes that gap:

search_rulebook_vector("structured address requirement", top_k=3)
{
  "query": "structured address requirement",
  "returned": 3,
  "results": [
    {
      "scheme": "HVPS+",
      "version": "2026",
      "clause": "structured-address-alignment",
      "title": "HVPS+ aligns with CBPR+ structured-address rule",
      "score": 0.702729,
      // ...
    }
  ]
}

Take the winning scheme/version/clause and pass it to cite_rulebook for the full citation.

It is not a neural embedding model. Retrieval is a deterministic lexical-vector cosine search: each clause and the query are hashed into a fixed 256-dimension term-frequency vector — whole words plus character 3- and 4-grams, so address matches addresses without a stemmer — bucketed with BLAKE2b rather than Python's salted hash, and ranked with sqlite-vec. That has three consequences worth knowing:

• The same query always returns the same ranking, in every process and every CI run.
• Nothing is downloaded and no network call happens at query time.
• It matches wording, not meaning. A query sharing no vocabulary with a clause will not find it, however related the concepts are.

Only the curated summaries are indexed — the same ones behind cite_rulebook. No external, copyrighted, or auth-gated rulebook text is stored or searched.

Installing it

sqlite-vec ships in an optional extra and is imported lazily, so the base install pulls in nothing:

python -m pip install 'camt053-mcp[vector]'

Without it the tool returns a plain error payload naming the extra rather than raising, so a client that calls it on a base install gets a usable message instead of a stack trace.

There is a second requirement that is easy to miss: your Python must be built with loadable SQLite extension support. The python.org macOS installers and several distribution packages ship it disabled, in which case sqlite-vec installs perfectly and still cannot load. The tool detects that and says so, naming the build flag (--enable-loadable-sqlite-extensions) to look for. Homebrew, uv and pyenv builds normally have it enabled.

See docs/vector-search.md for the retrieval design and its limits.

Using the tools

You can invoke the tools in-process — without a transport — straight through the FastMCP instance. This mirrors what an agent receives over stdio. The runnable version of this snippet lives in examples/mcp_tools.py.

import asyncio

from camt053_mcp.server import server

# A complete camt.053 statement with one entry returned AC04 (Closed Account).
statement_xml = """<?xml version="1.0" encoding="UTF-8"?>
<Document xmlns="urn:iso:std:iso:20022:tech:xsd:camt.053.001.14">
  <BkToCstmrStmt>
    <GrpHdr><MsgId>STMT-MSG-0001</MsgId><CreDtTm>2026-06-15T08:00:00</CreDtTm></GrpHdr>
    <Stmt>
      <Id>STMT-0001</Id><CreDtTm>2026-06-15T08:00:00</CreDtTm>
      <Acct><Id><IBAN>GB29NWBK60161331926819</IBAN></Id><Ccy>EUR</Ccy></Acct>
      <Bal><Tp><CdOrPrtry><Cd>CLBD</Cd></CdOrPrtry></Tp>
        <Amt Ccy="EUR">10000.00</Amt><CdtDbtInd>CRDT</CdtDbtInd>
        <Dt><Dt>2026-06-15</Dt></Dt></Bal>
      <Ntry>
        <NtryRef>NTRY-0001</NtryRef>
        <Amt Ccy="EUR">1500.00</Amt><CdtDbtInd>CRDT</CdtDbtInd>
        <Sts><Cd>BOOK</Cd></Sts>
        <NtryDtls><TxDtls>
          <RtrInf><Rsn><Cd>AC04</Cd></Rsn></RtrInf>
        </TxDtls></NtryDtls>
      </Ntry>
    </Stmt>
  </BkToCstmrStmt>
</Document>"""


async def main() -> None:
    async def call(name, args):
        result = await server.call_tool(name, args)
        # mcp 2.x returns a CallToolResult (read .content); 1.x
        # returns the content list, or a (content, meta) tuple.
        content = getattr(result, "content", None)
        if content is None:
            # mcp 2.x returns a CallToolResult (read .content); 1.x
            # returns the content list, or a (content, meta) tuple.
            content = getattr(result, "content", None)
            if content is None:
                content = result[0] if isinstance(result, tuple) else result
        return content[0].text if content else ""

    # Validate an identifier.
    print(await call("validate_identifier",
                     {"kind": "bic", "value": "NWBKGB2LXXX"}))
    # -> {"kind": "bic", "value": "NWBKGB2LXXX", "valid": true}

    # Page through the matching entries (paginated envelope).
    print(await call("filter_entries",
                     {"xml": statement_xml, "reason_code": "AC04",
                      "offset": 0, "limit": 5}))
    # -> {"total": 1, "offset": 0, "limit": 5, "entries": [...]}

    # Generate a validated reversing-entry document for the AC04 entries.
    xml = await call("generate_reversal",
                     {"xml": statement_xml, "reason_code": "AC04"})
    print(xml[:46])  # -> <?xml version="1.0" encoding="UTF-8"?> ...


asyncio.run(main())

Run it directly:

python examples/mcp_tools.py

The camt053 suite

camt053-mcp is part of a set of independently installable packages built around the camt053 library — pick whichever ones your stack needs:

Every tool here is a thin typed wrapper over camt053.services — the same facade the CLI, REST API, and LSP use — so all four interfaces behave identically.

When not to use camt053-mcp

• You have no MCP client. This server only makes sense paired with an MCP-aware host (Claude Desktop, the IDE plugins, an agent framework). For scripted / CI use, the camt053 CLI and REST API cover the same ground without the stdio protocol overhead.
• You need to run as a long-lived daemon without an MCP client. The server does run persistently over the streamable HTTP transport (--transport=http), but every consumer must still speak MCP JSON-RPC. For plain REST semantics, use the camt053 FastAPI service.
• You need streaming responses. Tool calls return whole values, not streams. Large statements are paginated through the existing list_entries(xml, offset, limit) envelope, not chunked over multiple responses.
• You need per-user OAuth flows brokered for you. The HTTP transport authenticates callers (OAuth 2.1 resource server with RFC 9728 metadata, or a static bearer token in dev mode) and scopes requests via the Camt053-Account tenant header, but it does not run an authorization server: bring your own IdP.
• You need to generate pain.001 outbound payment files. Out of scope; use pain001-mcp.

Development

camt053-mcp uses Poetry and mise.

git clone https://github.com/sebastienrousseau/camt053-mcp.git && cd camt053-mcp
mise install
poetry install
poetry shell

A Makefile orchestrates the quality gates (kept in lockstep with CI):

make check        # all gates (REQUIRED before commit)
make test         # pytest
make lint         # ruff + black
make type-check   # mypy --strict

Security

camt053-mcp is a thin wrapper — every tool delegates to camt053.services, where the defence-in-depth (defusedxml + xml_guard byte cap + DOCTYPE / ENTITY pre-flight) lives. Tools catch (ValueError, Camt053Error) and return an {"error": ...} envelope per the suite convention; they never propagate raw exceptions to the MCP client. Reporting practice, supported versions, and the full supply-chain posture are documented in SECURITY.md. Vulnerabilities go via GitHub Private Vulnerability Reporting, not public issues.

Documentation

• README.md — this file
• CHANGELOG.md — release notes
• SECURITY.md — disclosure + supported versions
• SUPPORT.md — how to get help
• MAINTAINERS.md — who can merge
• examples/ — runnable scripts
• glama.json — Glama directory manifest
• docs/iso20022-mcp-servers-compared.md — ISO 20022 MCP servers compared (2026)
• docs/BENCHMARKS.md — measured HTTP load benchmarks
• Glama listing: https://glama.ai/mcp/servers/sebastienrousseau/camt053-mcp

Related MCP Servers

Part of the ISO 20022 MCP Suite — open-source, Apache-2.0 licensed MCP servers for banking and financial-services AI agents:

MCP Registry

mcp-name: io.github.sebastienrousseau/camt053-mcp

License

Licensed under the Apache License, Version 2.0. Any contribution submitted for inclusion shall be licensed as above, without additional terms.

Contributing

Contributions are welcome — see the contributing instructions. Thanks to all contributors.

Acknowledgements

Built on the camt053 ISO 20022 Bank Statement library and the Model Context Protocol Python SDK.