AI should come to your data. Not the other way around.
MCPPlatform is the connector layer between Claude Desktop and the software your company already runs: accounting, GST, CRM, and sales intelligence. It is built by Recklabs in Delhi, and it is built on one conviction: the fastest way to make AI useful for a business is also the most private one.
01 / MISSION
Our mission
Indian businesses sit on years of structured financial truth, locked inside Zoho Books, TallyPrime, and the GST portal. Getting answers out of it means exports, pivot tables, and waiting on whoever owns the login.
We think the owner of a business should be able to ask, in plain English, "who owes me money and for how long," and get the real number in seconds. MCPPlatform exists to make that normal.
02 / BELIEF
What we believe
Privacy is not a feature tier here. It is the design constraint everything else was built around. We decided early that we would never warehouse customer financial data, because the best way to protect data we hold is to not hold it.
The orchestrator, the routing engine, and the databases run on your PC. Your credentials live in your local Claude Desktop config. If you stopped trusting us tomorrow, there would be nothing of yours on our side to ask back.
03 / ARCHITECTURE
How it works, in depth
When you ask Claude a question, the MCPPlatform engine on your machine takes over. It extracts the structured details, the month, the GSTIN, the ledger, and rewrites the rest into a reusable pattern. That pattern is embedded locally into a 768-dimension vector and matched against every route the platform has answered before, held in PostgreSQL with the pgvector extension. If a close match exists (cosine similarity ≥ 0.92), the route is reused in milliseconds, with no AI call at all.
For a genuinely new question, the engine shortlists the twelve most relevant tools and hands only those to Claude, which picks the exact tool and parameters. That resolved route is then saved back to the learning store, so the next similar question is an instant local hit. The platform compounds: more usage, faster answers, lower cost. Either way the tool runs locally and the result is shaped with decimal-safe arithmetic, stripped of personal data, and compressed to a short summary. Claude receives that summary and writes the answer, across 122 purpose-built tools, plus the Data Hub that coordinates them.
04 / INDIA
Why Indian businesses first
Most business software integrations are built for the American stack and ported east later, badly. We went the other way. TallyPrime support speaks to the Tally instance already running on your network and auto-detects your active company.
The GST connector fetches GSTR-2A and 2B and reconciles them against your purchases, because that is the actual monthly grind of a finance team here. Zoho Books gets 46 tools, not a token integration. And the local-first architecture lines up with the DPDP Act 2023, CERT-In Directions 2022, and IT Act Section 43A by default, which matters when your CA asks where the data goes.
05 / TRANSPARENCY
What leaves your device, stated plainly
Never leaves: raw financial data, ledger entries, invoices, amounts, customer and vendor names, API keys, OAuth tokens, database contents.
Can leave, only when the platform meets a new question: the text of that question, a few hundred tokens, sent for Claude to pick the right tool. Learned routes are stored locally as patterns with placeholders, never your org names, GSTINs, or amounts.
Also leaves, by definition: your account email and connector selection, which we need to compile your personalized installer. That is the whole list.
One honest caveat: Claude Desktop itself is Anthropic's product, and the summaries the platform hands to Claude are processed under Anthropic's terms. We keep those summaries small, shaped, and filtered for personal data, roughly 500 tokens per answer.
06 / STACK
The stack
See what your books can tell you.
Sign up free, pick your connectors, and ask your first question today.