TMCP (Model Context Protocol Server Implementation)

tmcp is a modern, composable, and flexible SDK designed for building Model Context Protocol (MCP) servers. It allows developers to build servers that enable LLMs (Large Language Models) to access additional context and act on behalf of users by standardizing communication between MCP clients (like claude-code, codex, or copilot) and MCP servers.

Key design principles of tmcp include:

• Flexibility: You choose your deployment target, validation library, and preferred transport.
• Modern Standards: It leverages standard patterns like the fetch API (Request/Response objects) and supports the Standard Schema specification, allowing users to use their preferred validation libraries (e.g., Valibot, Arktype, Effect, Zod).

For HTTP and SSE transports, tmcp implements a full OAuth 2.1 compliance framework. Key features include:

• PKCE Support: Proof Key for Code Exchange for secure authorization flows.
• Resource Indicators (RFC 8707): Support via the resource parameter for audience checking.
• Token Management: Support for authorization code flows, refresh tokens, and token validation.
• Client Registration: Support for dynamic client registration and multiple authentication methods.
• Access Control: Role-based access control (RBAC) using scopes and bearer token validation.
• CORS: Comprehensive Cross-Origin Resource Sharing configuration is integrated into the auth and transport layers.

MCP is built on JSON-RPC 2.0 over UTF-8 encoding. The protocol supports three main message types:

• Requests: Client-to-server or server-to-client operations with unique IDs, timeout, and cancellation support.
• Responses: Success results or error responses with standardized error codes.
• Notifications: One-way messages (e.g., real-time updates) that do not expect a response.

Supported Transports

• stdio Transport: Best for local processes and development. Uses standard input/output for efficient, low-overhead communication.
• Streamable HTTP Transport: Best for remote servers and web integrations. Uses POST requests and optionally Server-Sent Events (SSE). Requires HTTPS and origin validation.
• Custom Transports: Developers can implement custom transports (e.g., WebSockets, gRPC) as long as they implement JSON-RPC 2.0 message framing.

By default, SseTransport uses in-memory session management. For multi-server or serverless environments where sessions must be shared, use Redis-based session managers from @tmcp/session-manager-redis.

When to use Redis sessions:

• Multi-server deployments: When clients might connect to different instances.
• Serverless deployments: When instances are ephemeral (though note that serverless environments may kill SSE requests quickly; streaming-http is often preferred there).
• Load balancing: When requests are routed to different server instances.

import {
	RedisStreamSessionManager,
	RedisInfoSessionManager,
} from '@tmcp/session-manager-redis';

const transport = new SseTransport(server, {
	sessionManager: {
		streams: new RedisStreamSessionManager('redis://localhost:6379'),
		info: new RedisInfoSessionManager('redis://localhost:6379'),
	},
});

The 2025-06-18 version introduces enhanced capabilities for tool execution:

• Structured Tool Output: Tools now support output schema validation, which improves type safety and error handling when integrating with typed languages.
• Resource Links in Tool Results: Tools can now return resource links in their results, allowing for better integration and context sharing between tools and resources.

Unlike the official MCP SDK, tmcp is schema-agnostic. It uses an adapter to bridge your chosen validation library (like Zod, Valibot, or ArkType) with the Model Context Protocol.

Important Note on Adapters: An adapter is required if your server needs to:

• Accept structured inputs
• Produce structured outputs
• Request elicitations

If you pass adapter: undefined, the server will be extremely limited and cannot handle structured data. If you attempt to add a tool that uses a different validation library than the one configured in the McpServer adapter, the server will throw an error.

Tool results in tmcp are validated against a strict schema (CallToolResultSchema) to ensure compatibility with the MCP specification.

Supported Features:

• Content Types: Supports multiple content types including text, image, audio, and resource.
• Resource Links: Tools can return resource references in their results, allowing for seamless integration between tool execution and resource access.
• Validation: All tool outputs undergo JSON content validation and URI resolution for resource references.

Session management in TMCP is divided into three distinct concerns to handle different aspects of transport state:

1. Stream session managers: Handle long-lived streaming connections (like SSE/HTTP) and manage the fan-out of notifications to specific sessions.
2. Info session managers: Persist metadata that must survive across individual HTTP requests, such as client capabilities, clientInfo, requested log levels, and resource subscriptions.
3. Subscription managers: Route MCP 2026-07-28 per-request change notifications to long-lived subscriptions/listen streams.

Together, these managers handle session lifecycle (creation/deletion), session queries, message delivery, and tracking resource subscriptions.