# tally-aiagent — A Safety Gate for Agent Writes to Tally Prime Open-source Python library that stops an AI agent corrupting Tally's books. It refuses names the ledger doesn't hold and reads every voucher back after posting. Open source, by dave8172. Page: https://quirkyagents.com/wizard/projects/tally-aiagent/ ## Overview **In plain terms:** Tally Prime doesn't complain when software sends it a name it has never seen. It creates a new one and reports success. A person typing notices the unfamiliar entry appear in the dropdown. Software doesn't, and an AI agent least of all — it will confidently produce a plausible name it never checked. **tally-aiagent** is the seatbelt. Two rules, in code rather than in a manual: no name is ever written unless Tally already holds it, and every entry is read back out of Tally after posting and compared against what was intended. → [View on GitHub](https://github.com/dave8172/tally-aiagent) --- ## The failure, in one story Your system knows a supplier as `ACME SUPPLIES CO. LTD`. Tally knows the same supplier as `ACME SUPPLIES CO LTD` — no full stop after `CO`. You post a purchase bill. Nothing errors. Tally quietly opens a *second* supplier account under the slightly different name, books the bill against it, and returns `CREATED=1`. From that moment: - the supplier's balance is split across two accounts, so **neither one is right** - your payables report understates what you owe - statements you send that supplier don't reconcile - and nobody finds out until someone chases a payment, weeks later A trailing space does this. So does a comma, or an invisible character pasted out of Excel. The same thing happens to product names, and there it moves your stock valuation too. This is documented Tally behaviour, not a rare bug. **It is survivable when a human types and structurally unsafe when an agent does.** --- ## **What the library does about it** **Names come from Tally, not from the caller.** Your system supplies quantities, rates, dates and invoice numbers. Tally supplies names, read from its own master export. A name that can't be found there means the whole entry is refused and nothing is sent. **Ambiguity stops the line.** If `PSU-500` could mean either `PSU 500H` or `PSU 500S`, there is no safe answer. It shows both and refuses to guess. **"Success" is not proof.** After writing, the voucher is read back out of Tally and compared field by field. A mismatch is reported loudly instead of swallowed. --- ## Why an agent can't skip the check Writing is two steps, and the split is structural rather than a rule the agent is asked to follow. **Step 1 — prepare.** The agent describes the entry it wants. The library resolves every name against Tally, works out the money, and hands back a plain report. Nothing has been written yet. ``` party 'acme supplies pvt ltd' -> 'Acme Supplies Pvt Ltd' [matched] ledger 'Purchase Accounts' -> 'Purchase Accounts' [exact] item 'Widget A' -> 'Widget A' [exact] 10 x 125.50 = 1255.00 charge 'Freight Inward' -> 'Freight Inward' [exact] 450.00 totals goods=1255.00 charges=450.00 party=1705.00 ``` Left column: what the agent asked for. Right column: what Tally actually holds. Bottom: the money. That report is the thing a human reads. **Step 2 — approve, then post.** The `post` tool takes only a reference to a draft that already passed step 1. It has no parameter for describing a voucher at all, so there is no path to Tally that skips the report — the agent is not being trusted to behave, it simply has no other option. --- ## **No default ledger names. On purpose.** The library ships with an empty vocabulary. No suggested purchase ledger, no rounding account, no supplier-country default, no voucher numbering scheme. That's a deliberate cost. It was extracted from a production system that posts purchase bills into Tally every working day, and the tempting move was to carry that system's chart of accounts across as sensible defaults. A plausible default gets posted for months before anyone notices it was wrong, and the person who finds out is an accountant, not a developer. What leaks out of an internal tool is rarely the data. It's everything shaped around the data — vocabulary, ledger names, numbering conventions. --- ## Install ```bash pip install tally-aiagent ``` Library, CLI, and an MCP server, so an agent can use it directly. MIT licensed. → [Read the source](https://github.com/dave8172/tally-aiagent) --- ## Where it fits This is the same idea as [doceval](https://quirkyagents.com/wizard/projects/doceval/), one layer down. doceval measures whether AI output is *correct*. This one stops incorrect output from reaching a system that can't tell the difference. Both exist because "the model said it worked" is not evidence, and in accounting the gap between those two things is somebody's payables report.