Since our earlier piece on Claude and Tally Prime, the ecosystem around the Model Context Protocol (MCP) has grown fast — there are now half a dozen community MCP servers for Tally, and both Claude and ChatGPT can theoretically sit on the other end of the connection. But "theoretically" is doing a lot of work in that sentence: the two assistants get there through genuinely different architectures, with different practical and security trade-offs for a firm handling live client ledgers.
How Tally Prime talks to the outside world
Tally Prime doesn't expose a modern REST API. Everything routes through one of two long-standing mechanisms, both switched on from F1 → Settings → Advanced Configuration:
- The XML/HTTP gateway (port 9000 by default) — send it an "Export" XML request and it hands back ledgers, vouchers, or stock items; send an "Import" XML request and it can create vouchers or masters. This is read and write.
- The ODBC server (same port, separate toggle) — read-only SELECT queries against Tally's pseudo-tables (Vouchers, Ledgers, StockItems, etc.). No inserts or updates are possible over ODBC, which makes it the safer of the two when you only need an AI to look, not touch.
Tally Definition Language (TDL) sits underneath both — it's how you'd shape a custom report or add a menu button that fires an HTTP call, if the off-the-shelf export doesn't already give you what you need.
Connecting Claude: a closed loop on one machine
Claude Desktop and Claude Code connect to MCP servers defined in a local config file — claude_desktop_config.json, under %APPDATA%\Claude\ on Windows. A local server is launched as a subprocess and talks to Claude over stdio — standard input/output, not a network port. Nothing about that connection is reachable from outside your machine.
A handful of community MCP servers now target Tally Prime specifically — some read-only via ODBC, some with dozens of tools that can create or even delete vouchers and masters. Before pointing any of them at a live company, read the source: these are mostly small, unaudited open-source projects, not vendor-shipped software. For production use, a read-only ODBC-backed server is the safer starting point; reserve write-capable tools (voucher creation, master edits) for a sandboxed test company first.
Connecting ChatGPT: the catch
This is where the two assistants genuinely diverge. ChatGPT's MCP client — available through its Apps/Connectors framework — only speaks to remote servers over a public HTTPS endpoint (SSE or Streamable HTTP). It has no equivalent of Claude's local stdio launch. ChatGPT cannot reach localhost:9000 on your office PC directly, full stop.
| Assistant | Connection type | What that means for Tally |
|---|---|---|
| Claude Desktop / Code | Local, stdio subprocess | MCP server runs on the same PC as Tally. Nothing is exposed to the internet unless you deliberately choose a remote setup. |
| ChatGPT (Apps/Connectors) | Remote only, HTTPS (SSE/Streamable HTTP) | Requires a publicly reachable endpoint in front of Tally — a tunnel (e.g. Cloudflare Tunnel, ngrok) or a small hosted relay. Bigger attack surface by design. |
| ChatGPT — Custom GPT Actions (alternative) | Remote, OpenAPI-schema REST call | Doesn't need MCP at all — a lightweight REST wrapper around Tally's XML gateway, described with an OpenAPI schema, works today. Same "must be internet-reachable" constraint applies. |
If your practice is committed to Claude, the local-only path is genuinely simpler and safer to reason about. If you specifically need ChatGPT, budget time for setting up and securing a tunnel or relay — and treat that exposed endpoint with the same care you'd give any internet-facing system holding client financial data: authentication, IP allowlisting, and logging, at minimum.
What you can actually automate
- Daily reconciliation — pull the day's sales register and ask the assistant to flag entries that don't match the bank statement.
- Draft return summaries — query ledger groups tagged for VAT/tax and get a first-pass summary before you review and file.
- Anomaly flagging — scan a date range of vouchers for outlier amounts, duplicate entries, or narrations that don't match the ledger head.
- On-demand MIS — trial balance, P&L, or balance-sheet summaries without opening Tally and running the report manually.
The privacy question you have to answer first
Client ledgers are confidential by professional obligation, not just by preference. The moment any MCP relay, tunnel, or model API sits between Tally and the assistant, transaction data is leaving the local machine. Before connecting a live company, know exactly which hops your data passes through and who can see it at each one.
Beyond the basic exposure question, two risk classes are worth naming specifically: prompt injection — a malicious or malformed voucher narration is just text the model reads, and a sufficiently crafted one could try to redirect the assistant's next action — and unintended retention by whichever model provider or relay is in the loop. Neither risk is unique to Tally; both are well-documented general concerns with connecting LLMs to live business systems.
Our recommendation for a Nepali practice
- Start read-only. An ODBC-backed MCP server that can query but not write is enough for 90% of the useful automations above.
- Prefer Claude's local stdio setup over exposing Tally to the internet for ChatGPT, unless there's a specific reason you need ChatGPT.
- If you must use ChatGPT, put real authentication and an IP allowlist in front of the tunnel — don't rely on the endpoint being "hard to guess."
- Keep any write-capable tool (voucher creation, master edits, deletions) out of production until you've tested it thoroughly in a sandbox company.
- Review whichever MCP server's source code you're about to run — most of what's available today is community-maintained, not vendor-audited.
Want help wiring this up safely for your own books, or auditing an MCP setup someone else built for you? Get in touch.