@lucidbrain/sdk

Implements OAuth 2.1 with PKCE (Proof Key for Code Exchange) as a built-in MCP server capability, enabling secure delegated authentication without exposing credentials to client applications. Uses the authorization code flow with code challenge/verifier pairs to prevent authorization code interception attacks, particularly important for desktop and mobile MCP clients that cannot securely store client secrets.

Implements the WorkSpec v1.2 specification as a standardized schema pattern for defining MCP tool capabilities, inputs, outputs, and execution constraints. WorkSpec provides a declarative, JSON-based format for describing what a tool does, what parameters it accepts, what it returns, and how it should be invoked — enabling MCP clients to understand and compose tools without hardcoded knowledge of each tool's interface.

Provides a framework for creating and managing MCP server instances with built-in lifecycle hooks (initialization, connection, disconnection, shutdown). Handles the MCP protocol handshake, message routing, and resource cleanup automatically, abstracting away low-level protocol details so developers focus on tool implementation rather than transport layer concerns.

Enables dynamic registration of tools into an MCP server instance with automatic schema validation and capability advertisement. Tools are registered with WorkSpec schemas, and the server exposes a discovery mechanism allowing MCP clients to query available tools, their parameters, return types, and execution requirements without prior knowledge of the tool ecosystem.

Manages incoming MCP client connections with automatic protocol version negotiation, capability exchange, and connection state tracking. Handles the MCP handshake sequence (client hello, server hello, capability negotiation) and maintains connection state for message routing and authentication context propagation throughout the client session.

Executes registered tools with automatic parameter validation against WorkSpec schemas, type coercion, and structured error handling. Validates incoming tool invocation requests against the tool's schema, coerces parameters to expected types, executes the tool handler, and returns results or errors in a standardized format that MCP clients can parse and handle consistently.

Exposes server-side resources (files, databases, APIs) to MCP clients through a standardized resource interface, with support for streaming large resources and partial reads. Resources are defined with metadata (MIME type, size, last modified) and can be streamed to clients in chunks, enabling efficient access to large datasets without loading entire resources into memory.

Provides a prompt template system where MCP servers can define reusable prompt templates with variable substitution and execution context. Templates are registered with the server and can be invoked by MCP clients, enabling AI agents to access standardized prompts without hardcoding them in client code. Supports variable interpolation, conditional sections, and execution context passing.