Nearly every Nepali accounting practice runs on Tally Prime. Nearly every one of us has also spent the last year watching Claude and ChatGPT get dramatically better — while still typing every voucher by hand, exporting every trial balance to Excel, and re-keying every reconciliation. Anthropic's Model Context Protocol (MCP) is the piece that closes that gap: a small open standard that lets Claude read from and write to your Tally Prime directly. This guide explains what MCP is, how the Tally × Claude architecture actually works, what a CA can genuinely automate today, and where the guardrails have to sit.
Key takeaways
- MCP is Claude's tool interface. It is an open standard that lets Claude call functions on your machine — read a ledger, post a voucher, run a report — without you copy-pasting anything.
- Tally Prime already exposes an XML gateway and ODBC layer. An MCP server wraps those APIs into Claude-callable tools; the accounting engine itself is unchanged.
- The real productivity wins are ledger classification, bank reconciliation, VAT-return prep, and monthly-close narratives — not "AI does the whole audit."
- Give Claude read-only access first. Wire write tools behind explicit human approval — never let an agent post to a live production company file autonomously.
- For Nepali offices, the useful entry point is Claude Desktop (or Claude Code) + a local Tally MCP server, on the same LAN as the Tally installation.
What is MCP, in one paragraph
The Model Context Protocol is an open standard, published by Anthropic in late 2024, that lets an AI assistant call tools running on your own machine or network. Think of it as the "USB-C for AI" — a common shape so any client (Claude Desktop, Claude Code, other MCP-aware clients) can talk to any tool server (Postgres, GitHub, Google Drive, Slack, a filesystem, or — for us — Tally Prime). Instead of pasting data into a chat window, Claude asks the server: "please give me the sales register for Shrawan," or "please post this contra voucher." The server does the work in your existing system and returns the answer. You approve each destructive action.
The architecture — how Claude reaches Tally
Tally Prime already exposes two interfaces most accountants never touch:
- The XML gateway on port
9000— a Tally Definition Language (TDL) request/response endpoint. Send an XML request, get back a ledger, voucher list, trial balance, or a confirmation that a voucher was posted. - The ODBC layer on port
9000— read-only SQL-style queries over ledger, voucher, stock and company masters.
An MCP server is a small process — usually a Python or Node.js script — that:
- Speaks the MCP protocol on one side (so Claude can talk to it).
- Speaks Tally's XML/ODBC on the other side (so Tally can serve it).
- Exposes a small, deliberate set of tools —
list_ledgers,read_ledger,get_trial_balance,run_daybook,post_voucher,reconcile_bank. Each tool is a named function with typed inputs.
Claude Desktop (or Claude Code) is configured to launch that server on startup. When you ask Claude something like "what's the current outstanding for ABC Traders?", Claude picks the right tool, sends the request, receives Tally's answer, and phrases it back to you — with the underlying data still living in your Tally company file.
The Tally × Claude × MCP architecture
Claude Desktop → MCP server → Tally XML/ODBC. All three run locally; the Tally company file never touches the cloud.
Setup — the practical steps
- Enable Tally's ODBC/XML gateway. In Tally Prime: F1 (Help) → Settings → Connectivity → Client/Server configuration. Set the server as
Both, port9000. Confirm you can hithttp://localhost:9000in a browser and see Tally's response. - Install an MCP server for Tally. Several community servers are on GitHub (search "tally mcp server"). Pick one, install with
pipornpm, and read the source before running it — it will touch your books. - Install Claude Desktop from claude.ai/download (or use Claude Code from your terminal).
- Register the server in Claude Desktop's config file (
%APPDATA%\Claude\claude_desktop_config.jsonon Windows). Add an entry undermcpServerspointing at the launcher script and its args. Restart Claude Desktop. - Verify. Ask Claude something read-only first: "list all ledgers under Sundry Debtors." If it comes back with your actual customer list, the pipeline is up.
- Backup before writing. Take a copy of the company file. Only then experiment with write tools.
10 real automations a CA can run today
- Ledger classification. Import a CSV bank statement, ask Claude to map each transaction to the correct Tally ledger, and post as bank receipts/payments after review.
- Bank reconciliation. Feed the bank statement plus Tally's bank ledger; Claude produces the matched/unmatched pairs with a plain-English narrative.
- Monthly close narrative. "Compare this month's P&L to last month and explain the three biggest variances" — Claude runs the reports through the server and writes the note.
- VAT return prep. "Pull all VAT-registered sales and purchase transactions for Shrawan; group by tax rate and produce the numbers I need for the IRD portal." See our VAT return filing guide.
- TDS reconciliation. Match TDS deducted per Tally to Form 25-Ka data. Our TDS guide covers the underlying rates.
- Debtor ageing narratives. Trial-balance-to-note in one prompt: "list top-10 debtors above NPR 5 lakh and 90+ days, with call notes if any."
- Invoice-to-voucher. Drop a scanned bill in a folder; a filesystem MCP + Tally MCP pipeline extracts fields, drafts a purchase voucher, and shows it for approval before posting.
- Trial-balance queries in natural language. "What's the current balance of the ABC loan account and its interest paid this year?" — no more menu-diving.
- Audit sampling. "Give me a random sample of 25 vouchers above NPR 1 lakh from Ashadh, stratified by ledger." The MCP server draws from Tally; Claude formats the working paper.
- Client onboarding. Pointed at a new company file, Claude can produce a one-page profile: revenue trend, top customers/vendors, current ratios, open matters.
The wins are in the boring parts of the close — the parts where accountants already know what to do, but doing it burns hours. That is what MCP hands back.
Where the monthly-close hours actually go back — indicative Worklane data
Indicative internal figures — measured across three Worklane client closes after the pipeline stabilised. Actual savings depend on data quality and staff comfort.
Data privacy and safety
- Everything runs locally by default. The MCP server, Tally, and Claude Desktop all sit on your machine or LAN. Claude sends the model's reasoning to Anthropic; the data stays with you unless a tool explicitly returns it in an answer.
- Audit the server. The MCP server is the trust boundary. Read the source. Prefer a well-maintained open-source project or one you build yourself.
- Read-first, write-later. For the first month, only expose
read_*tools. Add write tools once the shape of what Claude asks for is predictable. - Human-in-the-loop for anything that touches books. Any voucher-posting tool should return a preview and require a second call to actually post — Claude Desktop already prompts before write actions.
- Log everything. Every tool call, every posted voucher, tagged with the user. If something goes wrong, you need the trail.
- Enterprise-tier data policy. Confirm your Claude plan's data-handling terms exclude your inputs from training. Consumer tiers may not.
Limits and honest failure modes
- Ledger names are messy in Nepal. Free-text ledger names full of typos and shorthand will confuse the classifier. Time spent cleaning masters up-front pays back tenfold.
- Company file access is single-user for writes. If Tally is single-user mode, the server needs exclusive access to post — coordinate with your team.
- Complex TDL customisations may need dedicated tools. If your firm has heavy TDL customisation, the MCP server may need matching handlers.
- The LLM will still hallucinate around gaps. If the server doesn't expose the field Claude needs, it may guess. Expose everything it should be able to see; refuse everything it shouldn't.
- Nepali dates. Convert BS ↔ AD at the server boundary — do not rely on Claude to do calendar arithmetic in production flows.
A hands-on walkthrough — Excel Kida on YouTube
The clearest video walkthrough for accountants in our region is on the Excel Kida YouTube channel — the same folks who put out excellent Excel and Tally training in Hindi/English. Their MCP series covers connecting Claude Desktop to Tally step by step, and it is well worth an hour on a Sunday afternoon before you attempt your first live setup.
- Channel: youtube.com/@excelkida
- Search their channel for "Tally MCP" or "Claude Tally automation" for the current playlist.
Treat their walkthrough as your visual companion; treat this guide as the CA's opinion on where to be conservative.
Frequently asked questions
Do I need Claude Pro or the API to use MCP?
Claude Desktop (free tier) supports MCP servers. Heavy usage — long tool chains, big files — benefits from a paid plan or the API, but the entry cost is zero.
Can Claude Code do this too?
Yes. Claude Code (the CLI) also speaks MCP and is the natural fit if you already work in a terminal — e.g., running the server as a background process on the accountant's workstation and driving it with a scripted workflow.
Does this replace my accountant?
No. It replaces the parts of the accountant's day that already felt like data entry. Judgment — provisions, classifications that matter, dealing with the client — still sits with a human. What you get back is time to do the judgment properly.
What about audit trail and ICAN expectations?
Every posted voucher must be traceable to a named human. Keep the server's log alongside the Tally audit log; treat AI-drafted vouchers exactly like a junior's data entry — reviewed and signed off by a senior before they hit the books.
Can I do the same with a different LLM?
MCP is open — any MCP-compatible client works. In practice, Claude Desktop and Claude Code are the smoothest today; ChatGPT and Gemini support similar patterns through their own tool interfaces. Our AI tools overview covers the wider landscape.
Getting your Tally + Claude setup done
Setting up a Tally MCP server safely, deciding which tools to expose read-only vs write, cleaning up ledger masters so the classifier is reliable, and training your team to review Claude's drafts before posting — this is exactly the kind of work a CA firm should own end-to-end.