Start free
ai-mcpConcept

Model Context Protocol

Model Context Protocol, or MCP, is an open standard that lets an AI assistant connect to an application's real functionality through a defined set of named tools, instead of only being able to chat about the application.

In short

The open standard behind "let my AI do my bookkeeping." Any MCP-compatible assistant — Claude, ChatGPT, Cursor, or others — can connect to BalanceMCP's nineteen tools through it, each one a narrow, specifically named action rather than open database access.

Also called: MCP

Model Context Protocol is an open standard that lets an AI assistant call an application's real functions through a defined set of named actions, called tools, instead of only being able to chat about the application or guess at an interface it can't actually see. It isn't specific to any one assistant — any MCP-compatible client can connect to any MCP-compatible server.

Bookkeeping specifically needs this kind of connection rather than "just ask the AI," because without it, an assistant only knows whatever's typed or pasted into a conversation — not an actual, current chart of accounts, real balances, or real transaction history. Ask an unconnected AI "what's my net income this month," and the answer is really an estimate based on what was described to it, not something verified against a real ledger.

BalanceMCP exposes nineteen specific tools over an authenticated endpoint secured by a bearer API key tied to one user's own books — import_statement, categorize_transactions, post_journal_entry, and sixteen others, each one a narrow, specifically named capability doing exactly what its name says and nothing more. Connecting an assistant this way is nothing like handing it a database password.

Every tool call runs under the exact same rules as anything else touching the books: the same tenant isolation, so a connected assistant genuinely cannot see or touch another user's books, and the same database-level guarantees covered elsewhere — entries must balance, the journal is append-only, nothing posts into a locked period. The AI isn't given a side door around any of that; it's given the same front door everything else uses.

What people get wrong

  • Assuming MCP is specific to one assistant — it's an open standard, and any compatible client can connect to any compatible server.
  • Assuming a connected assistant has open-ended database access rather than a fixed menu of narrowly scoped tools.
  • Assuming an AI's answer about "your books" is verified against real data unless it's actually connected through something like MCP — otherwise it's working from whatever was described to it.

Common questions

Is MCP specific to Claude?
No — it's an open standard. Any assistant that implements it, including others besides Claude, can connect to an MCP-compatible ledger the same way.
Is connecting an AI through MCP the same as giving it my login credentials?
No — it connects through a scoped API key tied to a fixed set of named tools, not raw database or account access. The assistant can only do what those specific tools allow.

Machine-readable: /api/knowledge/concept:model-context-protocol