MCP (Model Context Protocol)

Tip

MCP is one of the key building blocks of the AI agentic ecosystem. PromptFleet’s mcp_protocol crate works standalone — you don’t need the agent SDK to use it.

The Model Context Protocol is the open standard for connecting AI agents and LLM apps to external tools, data sources, and services. PromptFleet provides a full, WASM-compatible implementation.

What you get

• JSON-RPC 2.0 wire format with full request/response lifecycle
• Tool discovery (tools/list) and tool execution (tools/call)
• Memory operations for persistent context
• OAuth 2.1 Bearer token authentication
• HTTP+SSE and Streamable HTTP transports
• FastMCP compatible — interoperates with Python FastMCP and TypeScript MCP SDKs
• Client, server, and proxy modes via feature flags

Two ways to use MCP

Standalone (mcp_protocol crate)

Use pf_mcp_protocol directly when you want MCP in any Rust app — no agent SDK needed.

[dependencies]
mcp_protocol = { package = "pf_mcp_protocol", version = "0.1", features = ["client"] }

use mcp_protocol::{McpProtocolHandler, create_mcp_handler};

let handler = create_mcp_handler();
// Or customize:
// let handler = McpProtocolHandler::new()
//     .with_capabilities(ServerCapabilities::default())
//     .with_query_mode(mcp_protocol::QueryMode::Single);

Via the Agent SDK (mcp-client feature)

When building agents with agent_sdk, enable mcp-client to auto-discover and register MCP tools into the agent’s ToolRegistry. The SDK uses the correct backend automatically — NativeMcpBackend on native (via rmcp), WasmMcpBackend on WASM (via mcp_protocol + Spin HTTP).

[dependencies]
agent_sdk = { package = "pf_agent_sdk", version = "0.1", features = ["llm-engine", "mcp-client"] }

MCP tools become available to the LLM alongside local tools — the agent doesn’t need to know the difference.

Crate details

mcp_protocol Core MCP implementation — JSON-RPC 2.0, tool discovery, execution, auth, proxy. WASM + native.

Guides

• MCP Integration guide — step-by-step: connect external MCP servers, configure tool policies, wire into agents
• API reference — full type and module documentation

Feature flags

Feature	Purpose
client	MCP client module (included in default)
server	MCP server module (included in default)
proxy	Connect to external MCP servers outside Kubernetes (included in default)
sse-client	Direct SSE client — avoids proxy hop for performance
sse-server	Direct SSE server — accepts direct SSE connections