Tally MCP server: connect Tally to Claude or ChatGPT without risking your books

Picture a Surat textile trader on a Sunday evening. His CA has asked for the list of parties who are more than 60 days overdue, and the accountant who knows Tally inside out is at a wedding in Navsari.

He has Claude open on the laptop anyway. What he wants is to type "which parties owe me money for more than 60 days?" and get the answer from Tally. A Tally MCP server is the piece that makes that possible.

I think it's one of the more useful things you can connect to Tally right now. The catch is that the same connector can also change your books, and that's easy to miss in the setup steps. So that's where I'll start.

MCP (Model Context Protocol) is an open standard for connecting AI assistants like Claude and ChatGPT to outside tools and data. An MCP server is the connector for one system. A Tally MCP server sits between the AI and TallyPrime.

When you ask a question, the AI doesn't read your whole company file. It calls one of the server's tools, for example "trial balance for April to September", and TallyPrime sends back that report. The AI then reads the result and answers you.

So the AI only sees what it asks for, which beats pasting an Excel export into a chat. It's still your financial data going to an AI provider, though.

The one most people mean is tally-mcp-server on GitHub by Dhananjay Gokhale (github.com/dhananjay1405/tally-mcp-server). It's free under the MIT licence. ICAI's AI site lists it as a use case credited to CA Dhananjay Gokhale, with the line "Tally MCP server enables popular LLM to connect with Tally Prime ERP". CORAA wrote about that listing on 26 August 2026.

From its README, as of October 2026 (version 7.8):

• It works with TallyPrime Silver or Gold, and it needs Node.js on the machine.
• TallyPrime's XML port has to be on: F1 > Settings > Connectivity > Client/Server configuration, with TallyPrime as Server on port 9000.
• The README asks you to avoid the Educational version of TallyPrime because its date-range limits feed partial data to the AI and give wrong answers.
• It has around 20 tools: trial balance, profit and loss, balance sheet, stock summary, ledger and stock item balances, ledger statements, bills outstanding, chart of accounts, and a general query tool for anything else.
• It isn't read-only. Three tools write to Tally: one creates or updates ledgers, one deletes masters and one creates journal vouchers.
• Writes are allowed by default. Setting BLOCK_WRITE to 1 hides the write tools completely.
• Local mode runs it on the same PC as Tally. Cloud mode, for browser and mobile apps, is a web server setup the README itself calls "quite complicated".

The README also says: "Always backup your Company before instructing this tool." I'd take that seriously.

These are the kind of questions an owner or a CA would ask. For each one, I've noted what the server needs to touch. None of them need write access.

1. "Which parties have owed us money for more than 60 days?" Needs: bills outstanding (receivables).
2. "Which supplier bills fall due in the next 15 days?" Needs: bills outstanding (payables).
3. "How did gross profit this quarter compare with the same quarter last year?" Needs: profit and loss for two periods.
4. "What's our cash and bank balance today, account by account?" Needs: ledger balances for cash and bank ledgers.
5. "Show every entry in the Sharma Traders ledger since April." Needs: the ledger statement for one party.
6. "Which stock items haven't moved in 90 days?" Needs: stock summary and stock item statements.
7. "Which expense heads grew the most month on month?" Needs: profit and loss or ledger balances by month.
8. "Does the trial balance tally, and which ledgers have unusual balances, like a debit on a creditor?" Needs: trial balance and chart of accounts.
9. "List journal vouchers above Rs 1 lakh this month." Needs: the general query tool over vouchers.
10. "What's our balance sheet position at 30 September in simple words?" Needs: the balance sheet for that date.

Check the first few answers against Tally's own reports. An AI can add up wrongly or pick the wrong period, and you won't know unless you compare.

Go through this before anyone asks the first question. If you can't tick an item, stop there.

• Read-only. Set BLOCK_WRITE to 1 so the create, update and delete tools disappear. In Claude, also switch off any write tools in the connector settings, because Claude's Research mode can call connector tools without asking you each time.
• A copy of the company. Point it at a copy or backup of the company on a separate machine. If something goes wrong, you lose a copy.
• Query logs. Keep a record of which tools were called and when. CORAA suggests the same for CA firms, along with snapshots of the data used, so the audit file shows what was looked at.
• Where the data goes. Every answer means your Tally data travels to Anthropic or OpenAI. Know which account it runs under and who in the office can use it.
• Training. On Claude Free, Pro and Max, Anthropic uses your chats for training only if you allow it in the privacy settings. On its commercial products (Claude for Work and the API), Anthropic says it doesn't train on inputs or outputs by default. For ChatGPT, check your plan's data controls before you start.
• Passwords. In cloud mode the server has a login protected by a PASSWORD setting, and the default is literally "password". Change it to something long before the server goes online.
• Trusted servers only. Anthropic and OpenAI both warn that a malicious MCP server can trick the AI into doing things you didn't ask for. Use the GitHub version or one you've had checked, never a random fork.
• Approvals. Don't click "Allow always" until you've seen what each tool does.

Client data adds one more layer. CORAA puts it well: "MCP is a connection protocol, not a confidentiality waiver." Your duties to clients don't change because the tool is new.

Which ChatGPT plans can add a custom MCP server has changed more than once this year, so check OpenAI's help page for your plan before you start.

If your Tally runs on one office PC, start with the desktop route. Nothing has to face the internet.

I'd skip it, or wait, in these cases:

• You ask Tally the same three reports every week. A scheduled report on email or WhatsApp is cheaper and doesn't need anyone to type a question.
• Your masters are messy. Duplicate ledgers and wrong groups give wrong answers, faster. CORAA makes the same point: a connector doesn't fix poor master data.
• You need audit evidence. For working papers, CORAA says file-based exports are still the safer route in 2026.
• You run the Educational version. The README says it gives partial data.
• You want the AI to post entries straight into Tally. I'd wait on that. For bills, Tally's own Docs by Ira creates drafts that only reach your books once someone saves them as vouchers, and I've reviewed it at /blog/tally-ira-docs-by-ira-review/.
• Nobody in the office can look after a Node.js server and a port on the Tally machine.

Hand this part over. It assumes the desktop route with Claude on the same Windows PC as Tally.

1. Make a copy of the company and load it on the machine you'll use. Take a backup first, as the README asks.
2. In TallyPrime, enable the XML port: F1 > Settings > Connectivity > Client/Server configuration, TallyPrime as Server, port 9000.
3. Install Node.js, then get tally-mcp-server from GitHub (not the older v6 zip).
4. In Claude Desktop, install the extension or add the server to claude_desktop_config.json as the README shows.
5. Set BLOCK_WRITE to 1 and confirm the ledger, delete and voucher tools no longer appear in Claude's tool list.
6. Ask question 1 from the list above and match the answer with Tally's own outstanding report.
7. For browser or mobile access, follow the cloud-mode docs in the repo, change PASSWORD from the default, and decide who gets the URL.

If you'd rather not run this yourself, this is the kind of work we do at Dizilus: AI agents on top of the tools you already use, Tally included, read-only by default, with a person approving anything that changes a record. See how we work at /ai-agents/, or email tea@dizilus.com with the questions you'd want to ask your Tally.