This release is a pre-release and may not be stable for production use.
Finagent MCP — financial reports for AI assistants
Русский · Download 0.7.0rc2 · CI · MIT license
Read prepared financial reports with your AI assistant through the Model Context Protocol (MCP). Ask about profit and loss, cash flow, and account balances. Finagent exposes report rows, methodology, available months, and notes; the assistant explains the data within those limits.
0.7.0rc2 is a public prerelease. The repository and release downloads are available. The Python distribution name is open-finance-mcp; the server module is finance_mcp.server. PyPI and the official MCP Registry listing are not published yet.
Sources and requirements
| Source | Setup | Google approval required? |
|---|---|---|
| Synthetic demo | Install the package; no credentials or financial files | No |
| Excel (.xlsx) and CSV | Map prepared monthly reports and reconcile them against the source | No |
| PostgreSQL | Approved views, a read-only role, a source map, and reconciliation | No |
| Google Sheets | Google OAuth and a supported monthly report layout | Google data-access verification is pending |
PostgreSQL support does not imply MySQL or Microsoft SQL Server support. This is an adapter for prepared reports, not an automatic accounting engine for arbitrary transaction lists.
Your agent can inspect an approved source, prepare the mapping, run checks, and help configure the MCP client. Start with AGENTS.md and the guided setup (currently in Russian). The data owner confirms the meaning of the reports, monetary units, methodology, and whether a month is ready for analysis. The agent must reconcile all mapped rows; structural validation alone does not establish that amounts are correct.
Try the demo
Requirements: Python 3.11+, venv/pip, and network access for the initial dependency installation. Run these commands in a separate checkout:
git clone https://github.com/finuchetai/finagent-mcp.git
cd finagent-mcp
python3 preflight.py environment
python3 -m venv .venv
.venv/bin/python -m pip install .
.venv/bin/python -m finance_mcp.diagnostics stdio --timeout 15
.venv/bin/python -m unittest discover -s tests -p test_demo.py -v
On Windows, use python if python3 is unavailable and .venv\Scripts\python.exe in place of .venv/bin/python.
Configure a compatible local MCP client with the absolute path to that interpreter. The enclosing configuration format depends on the client:
{
"mcpServers": {
"finagent-demo": {
"command": "/absolute/path/to/finagent-mcp/.venv/bin/python",
"args": ["-m", "finance_mcp.server"]
}
}
}
Without --config, the server uses synthetic demo data. It does not connect to your accounts. Ask the assistant to read the passport and available months before requesting reports.
Example: “Show September 2026 profit, cash movement, and closing balance. Explain why profit differs from cash movement.” Demo control totals are 75,000 RUB profit, 80,000 RUB net cash movement, and 155,000 RUB closing balance. October 2026 is open and must be rejected. Demo dates are fixed for reproducibility.
Connect Excel, CSV, or PostgreSQL
Install the source extras into the same venv:
.venv/bin/python -m pip install '.[files,postgres]'
.venv/bin/python -m finance_mcp.server --config examples/lipa/csv.json --check
The example uses a fictional business. For client data, prepare and reconcile a source configuration, then append --config and its absolute JSON path to the MCP server arguments. Do not use an administrator DSN; PostgreSQL credentials must be restricted to approved read-only views. Keep credentials outside the published package. See source formats and permissions.
The default transport is local stdio. A remote client requires the owner's HTTPS server and an OAuth provider; this repository implements the resource server, not hosted login or a shared financial endpoint. See connection modes. Compatibility must be checked for the chosen client and account.
Google Sheets on Windows with Codex
Download finagent-google-codex-0.7.0rc2.zip, extract it into a separate folder, and run Start.cmd. You need installed Codex, Python 3.11+, and internet access. Follow the prompts for Google login, your existing spreadsheet link, report, month, currency, and monetary units. After successful setup, restart Codex and open a new chat to verify actual MCP tool calls.
As of 14 November 2026, Google branding is published and data-access verification is pending. Broad Google sign-in is not advertised as approved. If Google refuses access or shows an unverified-app warning, stop and contact the author with the error text; bypassing the warning is not an installation step. Do not share passwords, authorization codes, or tokens.
The Windows bundle includes the author's Desktop OAuth application configuration. A Git clone, the source ZIP, and a plain Python install do not include it. User grants stay in the Windows credential store; the bundle contains no user grant or financial spreadsheet. Supported layouts and setup prompts are documented in the Windows guide, currently in Russian.
Cash-flow reading was verified by an external tester in Codex on Windows, including a new chat after restart. The final Windows bundle was checked with the owner's existing Google permission. A new Google login and a new Codex chat for that final build remain separate checks. End-user acceptance of balances and other clients/platforms is not claimed.
Data boundaries
- Source data is read only. Installing the MCP does not authorize editing a workbook or database.
- Only approved closed months are accepted. A partial P&L is allowed with
complete=false; missing expenses or taxes are not invented. - Missing and not-applicable values remain distinct from zero. Totals and detail rows are not automatically added together.
- Reports identify currency, source, completeness, and timing. An unknown accounting date remains unknown; retrieval time does not replace it.
- The server holds a snapshot until restart. Google reports refresh automatically and fail on refresh errors. Stop the running process when revoking access.
Validation and support
CI for the published release passed stdio diagnostics and 98 tests on Python 3.11 and 3.13: 91 passed and 7 PostgreSQL runtime tests skipped on each version. Windows checks covered installation/reuse, 25 setup/demo/CSV/XLSX tests, and real Google/MCP reads with existing owner permission. Skipped tests are not PostgreSQL integration acceptance.
To disconnect, stop the MCP process and remove its client entry. For Codex use codex mcp remove <server-name>. Google permissions can also be revoked at Google account connections.
Report bugs through GitHub Issues without financial data, personal information, credentials, or tokens. See SECURITY.md, CONTRIBUTING.md, and CHANGELOG.md. These detailed documents are currently in Russian.
Self-hosting does not require a Finagent service account. For assistance with setup or management accounting, contact the Finagent team.
Metadata
Release files for open-finance-mcp 0.7.0rc2
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| open_finance_mcp-0.7.0rc2.tar.gz | 179.0 kB | With provenance | Details for open_finance_mcp-0.7.0rc2.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| open_finance_mcp-0.7.0rc2-py3-none-any.whl | Python 3 | none | any | Details for open_finance_mcp-0.7.0rc2-py3-none-any.whl |
Total release size: 277.5 kB
Release files / open_finance_mcp-0.7.0rc2.tar.gz
| Download URL | open_finance_mcp-0.7.0rc2.tar.gz |
|---|---|
| Size | 179.0 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
75004aae09080abae200bbfc0e9dc3e081a153a2a0a3dc9c759dafcc613021f2
|
|
BLAKE2b-256 checksum How to use checksums |
fddc0b28e11063bbd30ed94058f37ad713b4bb9265dad72e1982baedc89d7a92
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 14, 2026.
Transparency logRelease files / open_finance_mcp-0.7.0rc2-py3-none-any.whl
| Download URL | open_finance_mcp-0.7.0rc2-py3-none-any.whl |
|---|---|
| Size | 98.6 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
efbbd01503d6f4f6c942addec2b8e0a21a98b8852a2fa9e60011ce6cf378633c
|
|
BLAKE2b-256 checksum How to use checksums |
8293ec4c3e9db7b448f5cdbe5c1c3fc23d023c1f46d0341fc766825ed8bbcb84
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 14, 2026.
Transparency logSecurity for open-finance-mcp 0.7.0rc2
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
Full provenance
Both files in this release have been uploaded with provenance metadata.
Attestation for open_finance_mcp-0.7.0rc2.tar.gz
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 14, 2026.
Transparency logA note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.