Consume HTTP and MCP interfaces

Call the same Python calculation from an HTTP application and a tool client. Verify the actual public names, response and invalid input handling.

Before you begin

Generate the quote example or Python compiler example. Install the emitted transport requirements in a separate service environment. Run only trusted bundled code and dependencies. The example contains complete http_client.py and mcp_client.py files.

Call HTTP

Start the quote REST bundle using the local lifecycle procedure. From the extracted example:

python http_client.py http://127.0.0.1:8000 order.json

The client sends POST /capabilities/api.quote, validates the JSON response and prints it. HTTP 422 is input-shape failure; the example's ok:false response is a business refusal. An unexpected function exception becomes a generic 500, not an invitation to send source exception messages to clients.

Call a generated MCP tool

The MCP client initializes a stdio session and checks the real tool inventory:

../runtime/bin/python mcp_client.py ../runtime/bin/python ../bundles/mcp/server.py order.json

The server is started by the client and closed when its context exits. There is no separate background stdio server to manage. result.is_error signals a tool error; structured_content contains the business object. Scalar results instead use {"result": value}. The quote's object remains an object.

MCP Tool names include modules for repository bundles (api.quote), whereas a single-source generated tool uses the selected symbol (plan_route). Read the emitted tool list instead of assuming a name.

Distinguish tool execution from assistant decisions

These calls use the official Python MCP SDK and need no AI account. An assistant may choose a tool and arguments, but the calculation is performed by your function through MCP. The route example makes that distinction visible without a simulated model response.

The optional analysis MCP plugin supplies Apizr repository-analysis tools, not your business functions. Configure it separately if your integration needs exploratory analysis.

Add access controls in your deployment

The tutorials bind HTTP locally and use trusted code. Before remote access, your platform must provide authentication, authorization and lifecycle controls. Generated contracts and a successful local call do not implement organization permissions. Continue with operations.