# How to add AI agents to TallyPrime > Source: https://netray.co/integrations/tally-ai-agents > Publisher: Netray (netray.co), AI and integration engineering for ERP and manufacturing systems. > Section: Integrations ## Short answer TallyPrime has no agent framework of its own, but it exposes an HTTP interface on the machine running Tally (XML, and JSON in newer releases), plus ODBC and TDL for custom collections. An agent sits outside Tally, calls those interfaces as typed tools, reads freely, and proposes vouchers that a person approves before they are imported. ## System TallyPrime ## Capability ai-agents ## Intro Most manufacturers on TallyPrime run it as a single-company, single-machine or small-LAN installation. The books are accurate, but the work around them is manual: chasing receivables, matching supplier bills to purchase vouchers, reconciling stock after a production run, and preparing GST workings. These are exactly the repetitive, rules-heavy tasks an AI agent can take on, as long as it is kept away from direct, unreviewed writes to the books. The difficulty is that Tally is not a server-first product. The company data lives in local files, the HTTP interface works only while the company is loaded, and there is no per-user API key model. An agent design has to respect that: a thin service next to Tally, a defined list of tools, and an approval step that a bookkeeper or accountant actually sees. This page covers how to wire that up, which agent jobs are worth building first, and where a plain TDL report or a scheduled export is enough and an agent is overkill. ## Use Cases ### Receivables chasing **Description:** An agent reads outstanding bills by party and ageing, drafts a reminder per customer with the invoice numbers and due dates, and queues it for approval. It never sends without a person confirming the list. ### Supplier bill matching **Description:** The agent compares a scanned or emailed supplier bill with purchase vouchers and GRN style stock entries, flags rate and quantity differences, and prepares a purchase voucher draft for review. ### Stock and production variance review **Description:** After a manufacturing journal or stock journal is posted, the agent compares consumed against standard BOM quantities, lists items with unusual wastage, and writes a short exception note for the owner. ### GST workings assistant **Description:** The agent pulls the GST reports for the month, flags vouchers with missing GSTIN, HSN or place of supply issues, and lists fixes for the accountant instead of changing data itself. ### Month end close checklist **Description:** A scheduled agent checks for unposted drafts, negative stock, bank ledgers with uncleared entries and large cash balances, then sends a summary to the finance head. ## System Surface HTTP server inside TallyPrime: Tally can listen on a configurable port (9000 is the commonly seen default) and accept XML request envelopes for export and import. Newer releases also accept JSON; confirm against your TallyPrime release notes. XML envelope structure: a HEADER with request type (Export or Import) and the object or collection ID, a BODY with static variables such as the current company and export format, and for imports a TALLYMESSAGE block holding the VOUCHER, LEDGER or STOCKITEM data. TDL (Tally Definition Language): custom collections and reports can be added through a TDL file so that the agent fetches exactly the fields it needs, for example outstanding bills with custom UDF columns, instead of parsing a full day book. ODBC: TallyPrime can act as an ODBC data source for read-only tabular access to ledgers, vouchers and stock items, handy for agents that prefer SQL style tools. Field names and available tables vary by release. Company must be loaded: the interface answers for companies that are open in the running Tally instance, so the agent service needs a way to detect or handle a closed company, and multi-company setups need the company name passed explicitly. ## How It Works ### Enable and secure the Tally interface **Detail:** Turn on the client/server or ODBC settings in TallyPrime, bind to a known port, and keep it reachable only from the agent host through the firewall. Do not expose the port to the wider network. ### Build a small tool layer **Detail:** Write a service that wraps the XML or JSON calls into typed functions such as get_outstanding(party), get_stock_summary(item), list_vouchers(type, from, to) and draft_voucher(payload). The model never builds raw XML itself. ### Add TDL collections for the heavy reads **Detail:** For reports that are slow or noisy in the default export, load a TDL file that defines a lean collection. Test its output on a copy of the company before pointing the agent at the live one. ### Choose the model deployment **Detail:** Either a hosted model with only the minimum fields sent, or an open-weight model served locally with vLLM or Ollama if party and pricing data must not leave the premises. ### Put approvals in front of imports **Detail:** Agent output for writes is a draft: a rendered voucher plus the XML to be imported. A named person approves it in a simple web page, and only then the service performs the import call and records the result. ### Log and replay **Detail:** Store every tool call, the exact request sent to Tally and the response, so an accountant can see why a voucher exists and the team can replay failures after a Tally upgrade. ## Example Flows ### Weekly overdue receivables sweep **Trigger:** Monday 9:00 scheduled run. **Steps:** - Agent calls the outstanding receivables collection and groups bills by party and ageing bucket. - It checks the party master for contact details and credit terms. - It drafts one reminder per party, citing bill numbers and amounts. - The accounts head reviews the batch in a web page and ticks the ones to send. **Result:** Reminders go out the same morning with an audit trail, and nothing is written into Tally. ### Supplier bill to purchase voucher draft **Trigger:** A supplier invoice arrives in the accounts mailbox. **Steps:** - Agent extracts supplier, bill number, items, rates and GST from the document. - It looks up the supplier ledger, stock items and open purchase orders if tracked. - It highlights mismatches in rate or quantity. - It renders a draft purchase voucher and the import XML for review. - On approval, the service imports the voucher and stores the returned voucher number. **Result:** Entry time drops and every posted voucher links back to its source document. ## On Prem TallyPrime is usually on-prem already, so the agent can be too. Run the tool service on the same machine or a small server on the LAN, serve an open-weight model on a local GPU box if the data must stay inside the building, and keep the Tally port closed to everything except that host. No cloud connection is required for the core flows. The practical constraint is availability: if the Tally machine is off or the company is closed, the agent has nothing to read, so schedule runs inside business hours or keep a nightly export as a fallback data source. ## Security - Tally has no fine grained API roles, so the tool layer must enforce permissions: expose read tools by default and a single draft tool for writes. - Never expose the Tally HTTP port beyond the agent host; treat it as an unauthenticated admin interface. - Redact bank account numbers, PAN and similar identifiers before sending text to any hosted model. - Keep a signed audit record of who approved each import and the exact payload that was sent. ## Options ### Scheduled exports plus a general assistant **Fit:** Small firms that want summaries and reminders but not live access. **Effort:** Days, low risk, data is a few hours old. ### In-house tool service over XML or ODBC **Fit:** Teams with a developer who knows Tally XML and TDL. **Effort:** Several weeks including testing on company copies. ### Partner build with Netray and a Tally partner **Fit:** Firms wanting on-prem models, approvals and audit trail without building the plumbing. **Effort:** A scoped pilot first, then agent by agent. ## Pitfalls - Importing vouchers with a ledger name that does not exactly match the master creates a new ledger or fails, so validate names against masters first. - A company that is closed or a Tally screen sitting on a modal dialog can stall requests; add timeouts and clear errors. - Large day book exports are slow and verbose; use TDL collections or date windows instead of whole-year pulls. - Behaviour differs between Tally.ERP 9 and TallyPrime, and between releases; pin and test against your exact version. - Letting an agent post directly to books makes errors hard to unwind because Tally edits are visible only through alterations; keep approvals mandatory. ## Faq ### Can an AI agent post vouchers into TallyPrime directly? **Answer:** Technically yes, through the XML or JSON import request. Practically, it should only post after a named person approves a draft. Voucher imports change the books and may trigger GST consequences, so keep the agent in a propose-and-approve role until you have months of clean drafts to justify more autonomy. ### Do I need the cloud to run this? **Answer:** No. The agent service and an open-weight model can run on the same LAN as TallyPrime. Cloud models are optional and are best limited to non-sensitive text such as reminder wording. ### Does this work with Tally.ERP 9? **Answer:** Much of the XML interface is similar, but field names, JSON support and connected services differ. Treat TallyPrime and Tally.ERP 9 as separate targets and test your specific release before building tools. ### Is an agent better than a TDL report? **Answer:** For fixed reports, a TDL customisation is simpler and cheaper. An agent earns its place where the work involves documents, judgement, drafting messages or combining Tally with email and spreadsheets. ### What happens if the Tally machine is off? **Answer:** The tools fail with a clear error and the agent reports it instead of guessing. Many teams keep a nightly export to a database so read-only questions still work when Tally is unavailable. ## Search terms tally ai agents, tallyprime ai agent, ai agent for tally, tally automation with ai, tally xml api ai, tally erp ai automation, add ai to tallyprime, tally voucher automation ai, on-prem ai for tally --- Talk to an engineer: https://netray.co/contact?intent=erp_ai | Machine catalog: https://netray.co/agent/catalog.json