Start free
ai-mcpConcept

MCP Client

An MCP client is the AI assistant — any Model-Context-Protocol-compatible assistant — that calls an MCP server's tools on a user's behalf.

In short

The AI side of the connection. Any MCP-compatible assistant can act as a client to BalanceMCP's server, calling its nineteen tools under the exact same rules — balanced entries, no editing history, no bypassing tenant isolation — regardless of which assistant it is.

Also called: connected AI assistant

An MCP client is the AI assistant that initiates tool calls against an MCP server — the half of the connection sitting on the user's side, translating a plain-language request like "categorize my June transactions" into the specific tool calls that actually do the work.

Any assistant that implements the protocol can act as a client to BalanceMCP's server; it isn't limited to one specific product. What makes a client trustworthy isn't its own judgment or carefulness — it's that every tool it calls runs under the exact same database-level rules regardless of which client is calling: balanced entries only, an append-only journal, no posting into a locked period, and tenant isolation that keeps one user's books invisible to another's session entirely.

A client only has access to what the tools it's given actually allow — there's no way for a client to reach past a tool's own scope into some broader capability. Reconciling an account through a client still requires calling reconcile_account with the right arguments; there's no shortcut around the interface the server exposes.

The API key a client uses is what ties every one of its tool calls back to a specific user's books — a client with a valid key can do everything that key's tools allow, which is exactly why the key itself, not the client's own behavior, is the thing worth protecting carefully.

What people get wrong

  • Assuming only one specific assistant can act as an MCP client — any protocol-compatible assistant can.
  • Assuming the client's own carefulness, rather than the server's database-level rules, is what keeps the books safe.
  • Assuming a client can reach capability beyond what its available tools expose — it's limited to exactly the named actions it's been given.

Common questions

What actually makes an AI client safe to use with real books?
Not the client's own judgment — the database-level rules every tool call runs under, regardless of which client is calling: balanced entries, an append-only journal, and tenant isolation.
Can any AI assistant act as a client, or only specific ones?
Any assistant implementing Model Context Protocol can act as a client — it isn't limited to one specific product.

Machine-readable: /api/knowledge/concept:mcp-client