Workflow · AI agent builders

Knot your API to any AI agent

You want Claude / Cursor / Copilot to call your internal services with the right shape of arguments, not a free-text prompt you have to keep fixing.

Bowire’s MCP adapter turns every discovered method into an MCP tool with JSON Schema input, validates arguments before invocation, and streams results back over SSE or stdio — zero hand-written adapter code.

One flag, every method becomes a tool

No manual tool definitions. Bowire enumerates your services, maps each method’s input schema to JSON Schema, and exposes the lot via the MCP streamable-HTTP transport.

Enable the adapter

--enable-mcp-adapter on the CLI, or .WithMcpAdapter() chained onto MapBowire() in embedded mode. That’s the whole opt-in.

# Standalone, against any server
bowire --url https://api.example.com --enable-mcp-adapter

# Then in Claude Desktop config:
# "my-api": {
#   "url": "http://localhost:5000/bowire/mcp"
# }

Bowire hosts the MCP endpoint at {prefix}/mcp, responds to initialize, tools/list, tools/call, and ping. Every unary method Bowire discovers — across gRPC, REST, SignalR, GraphQL, any protocol plugin — becomes a typed tool.

MCP plugin docs →

Bowire is also an MCP client

Point Bowire at someone else’s MCP server and browse their tools / resources / prompts in the same sidebar — typed forms, response viewer, history.

Inspect tools published by other agents

Your agent doesn’t work? Bowire becomes a debug probe for the same MCP server it’s hitting.

bowire --url http://localhost:5000/mcp runs the discovery pass against an MCP server, surfaces tools, resources, and prompts as separate service categories in the sidebar, maps each tool’s inputSchema to Bowire’s form UI, and dispatches invocations through the same JSON-RPC client an agent would use.

Streaming tools (SSE transport, chunked responses) render in the same Wireshark-style message view as gRPC streams.

Streaming view in detail →

MCP streaming tool — response frames rendered in the Wireshark-style pane

Debug the assistant’s side of things

Record the calls the agent is making and compare them to what the agent claims it did. Assertions on the response shape catch tool drift early.

Record agent sessions for reproducibility

Turn on recording while the agent is running, capture every tool invocation, replay against a test env to confirm behaviour.

Every tool call the adapter handles is a regular Bowire request — same recording, same environment substitution, same assertions. Great for post-mortems when an agent “hallucinated” a result that actually came from a call you didn’t expect.

Recording in detail →

Recording manager — sequence of captured tool calls

Ship typed tools to your agent

Flip one flag, get a full MCP tool catalogue. Debug with the same UI you use to invoke the services by hand.

Go to the MCP docs