TallyPrime MCP

Read-only MCP server that lets AI assistants (Claude Desktop, Cursor, and other MCP clients) query a TallyPrime installation running on your own computer.

Everything stays local: this server talks to TallyPrime's HTTP-XML gateway on 127.0.0.1, and it never sends your accounting data to any server operated by this project.

Status: v0.1.0, early alpha. The XML request/response shapes are built from TallyPrime's own published integration documentation, but have not yet been verified against a real, running TallyPrime installation in this project's own development environment. See Testing below for exactly what has and hasn't been confirmed, and please open an issue with your TallyPrime version if something doesn't match.

Features

• 13 read-only tools covering companies, ledgers, vouchers, trial balance, P&L, balance sheet, receivables, payables, and stock/inventory.

• Nothing can write to Tally. There is no execute_tally_xml-style escape hatch, and no tool can create, edit, or delete anything — see Security.

• Runs entirely on your machine. No cloud account, no API keys, no external server.

• Clean, typed, consistent JSON — never raw Tally XML tag soup — with amounts as numbers and dates as YYYY-MM-DD.

• Configurable, not hard-coded. Host, port, timeouts, default company, and limits are all environment-variable driven.

Related MCP server: Tally MCP Server

Architecture

flowchart LR
    A[AI Client<br/>Claude Desktop / Cursor / etc.] -->|MCP protocol, stdio| B[TallyPrime MCP Server]
    B -->|Local HTTP + XML| C[TallyPrime<br/>HTTP-XML Gateway]
    C --> D[(Local Accounting Data)]

The server is organized as:

tallyprime_mcp/
├── tally/       # HTTP transport, XML request building/parsing, custom exceptions
├── models/      # Pydantic response shapes (Company, Ledger, Voucher, reports, ...)
├── services/    # Maps raw Tally data onto the models above
├── security/    # Write-permission gate (deny-by-default), audit logging
├── mcp/         # MCP tool / resource / prompt registration
├── server.py    # Wires it all together into a FastMCP app
└── cli.py       # `tallyprime-mcp` command-line entry point

Requirements

• TallyPrime, running on Windows (the normal deployment target — Linux/macOS work for development of this server, but TallyPrime itself is a Windows product), with the HTTP-XML gateway enabled and a company loaded.

• Python 3.11+

• This server and TallyPrime should run on the same machine, or at least the same trusted local network — see Security before doing the latter.

Installation

If you've never used Python before: see docs/installation.md for a step-by-step, no-assumptions walkthrough.

If you're comfortable with Python:

pip install tallyprime-mcp

or, from source:

git clone https://github.com/cavishal04/tallyprime-mcp.git
cd tallyprime-mcp
pip install -e ".[dev]"

TallyPrime configuration

You need TallyPrime's HTTP-XML gateway switched on and listening (default 9000) with the company you want to query open. Exact menu paths vary by TallyPrime release — see docs/tally-setup.md for current instructions and screenshots-in-words.

Running the server

tallyprime-mcp

This starts the MCP server over stdio, which is what Claude Desktop, Cursor, and similar clients expect for a locally-run tool. You won't see much on screen — that's expected; it's waiting for an MCP client to connect.

Check connectivity first if you're not sure Tally is reachable:

tallyprime-mcp test-connection

Configuration

All configuration is via environment variables (or a .env file in the working directory — see .env.example):

Full list: docs/configuration.md.

MCP client configuration

• Claude Desktop — examples/claude

• Cursor — examples/cursor

• Any other MCP client — examples/generic

Full walkthrough: docs/mcp-clients.md.

Available tools

Full parameter reference: docs/tools.md (this is also where the trial-balance/P&L/balance-sheet design tradeoffs are explained — read it before relying on those three for anything important).

Example AI interaction

User: "Show me the trial balance for April 2026."

AI: calls get_trial_balance(from_date="2026-04-01", to_date="2026-04-30")

TallyPrime MCP: queries the local TallyPrime installation, returns structured ledger balances

AI: "Here's the trial balance — total debits and credits both come to ₹18,42,000. The largest debit balances are HDFC Bank (₹6,10,000) and..."

Security

Read SECURITY.md for the full threat model. In short:

• Every tool is read-only. No tool can create, modify, or delete Tally data.

• There is no generic XML-execution tool.

• The server binds to 127.0.0.1 by default and is never meant to be exposed to the internet.

• All input is validated (dates, limits, request size) before being sent to Tally.

• Nothing is logged that shouldn't be (see logging_config.py's redaction filter).

Privacy

• TallyPrime MCP runs entirely on your computer. There is no cloud backend.

• This project's maintainers do not receive, store, or have access to your accounting data, in any form.

• This is distinct from your AI client's own privacy policy. Once TallyPrime MCP hands data to your AI assistant (Claude Desktop, Cursor, etc.) inside your local MCP session, that data is subject to that product's privacy practices, not this project's — check your AI client's documentation for how it handles data sent to it.

Development

git clone https://github.com/cavishal04/tallyprime-mcp.git
cd tallyprime-mcp
pip install -e ".[dev]"
ruff check src tests
pytest

See docs/development.md for the full contributor setup.

Testing

• Unit tests (98 at last count) run against hand-built mock XML fixtures shaped to match TallyPrime's documented request/response patterns — they verify this project's own request-building, parsing, and business logic, but do not verify that a real TallyPrime installation responds exactly this way.

• Integration tests (tests/integration/) are written to run against a real, locally-running TallyPrime instance, but are skipped by default (including in CI) because no TallyPrime installation was available in this project's development environment. Run them yourself with TALLYPRIME_MCP_RUN_INTEGRATION=1 pytest tests/integration -v once you have TallyPrime open locally, and please report back via an issue with what did/didn't work for your version.

If you do have TallyPrime available and something in tools.md's "verify against a live instance" notes turns out wrong, a bug report (or PR) with the corrected field name is one of the most valuable contributions you can make right now.

Contributing

Contributions are very welcome — see CONTRIBUTING.md, especially if you have a real TallyPrime installation to test against. CODE_OF_CONDUCT.md applies to all project spaces.

Roadmap

• Verify all XML field names against a real TallyPrime installation (multiple versions)

• Write operations (create_payment, create_receipt, create_sales_voucher, ...), gated behind the confirmation workflow already scaffolded in security/confirmation.py — not started, and will not ship without explicit user confirmation per write

• Cost centre / godown-aware reporting

• GST-specific reports

• Optional local caching layer for large ledgers

License

Apache License 2.0 — see LICENSE. Apache-2.0 was chosen (over MIT) for its explicit patent grant, which feels appropriate for a project that talks to commercial accounting software and may attract contributions from companies as well as individuals.

This project contains no TallyPrime source code and no copied Tally Solutions documentation — only original code written against publicly available integration documentation.

TallyPrime is a product of Tally Solutions Pvt. Ltd. This is an independent, community open-source project and is not affiliated with, endorsed by, or sponsored by Tally Solutions.