@z_ai/mcp-server

Implements Model Context Protocol server that bridges MCP clients (Claude Desktop, IDEs, agents) to Z.AI's backend API infrastructure. Uses stdio/SSE transport to expose Z.AI's language models, vision models, and tool capabilities through standardized MCP protocol, abstracting away Z.AI API authentication (Bearer token), endpoint routing, and request/response marshaling. Handles protocol negotiation, capability advertisement, and bidirectional message passing between MCP client and Z.AI backend.

Exposes Z.AI's language model family (GLM-5.1, GLM-5, GLM-5-Turbo, GLM-4.7, GLM-4.6, GLM-4.5, GLM-4-32B-0414-128K) through MCP tool interface, routing requests to appropriate model based on capability requirements (context window, latency, cost). Implements model selection logic that abstracts model-specific parameters, token limits, and performance characteristics. Supports streaming and batch inference modes with configurable temperature, top-p, and other generation parameters.

Implements Bearer token authentication for Z.AI API access, accepting API keys from Z.AI Open Platform and converting them to Bearer tokens for API requests. Handles token lifecycle (generation, refresh if applicable, expiration), secure storage (environment variables or secure config), and per-request token injection into Authorization headers. Implements error handling for invalid/expired tokens with clear error messages.

Implements MCP protocol capability advertisement, informing clients of available models, tools, and resources exposed by the server. Uses MCP protocol initialization handshake to exchange supported capabilities, protocol version, and implementation details. Enables clients to discover available models (GLM-5.1, GLM-5V-Turbo, CogView-4, etc.) and tools (web search, function calling, etc.) without hardcoding assumptions.

Exposes Z.AI's vision model family (GLM-5V-Turbo, GLM-4.6V, GLM-4.5V) and specialized models (GLM-OCR for document extraction, AutoGLM-Phone-Multilingual for mobile UI understanding) through MCP tool interface. Accepts image inputs (base64, URL, or file path) and processes them with vision-specific models, returning structured analysis (object detection, text extraction, scene understanding, OCR results). Implements image preprocessing (resizing, format conversion) and model-specific input validation.

Exposes Z.AI's image generation model (CogView-4) through MCP tool interface, accepting text prompts and optional style parameters to generate images. Implements prompt processing, style embedding, and image encoding (base64 or URL return format). Supports iterative refinement through prompt modification without explicit inpainting, leveraging CogView-4's prompt understanding for style consistency.

Exposes Z.AI's video generation models (CogVideoX-3, Vidu Q1, Vidu 2) through MCP tool interface, accepting text prompts or image+text inputs to generate short videos. Implements video encoding, streaming output, and asynchronous generation handling (polling or webhook-based completion notification). Supports different video quality/length tradeoffs across model variants.

Exposes Z.AI's automatic speech recognition model (GLM-ASR-2512) through MCP tool interface, accepting audio input (file, URL, or stream) and returning transcribed text with optional speaker identification and timestamp metadata. Implements audio format detection, preprocessing (resampling, normalization), and streaming transcription for long audio files.

Implements function calling capability through MCP tool interface, allowing Z.AI language models to invoke external functions/tools based on semantic understanding of task requirements. Uses JSON Schema to define function signatures, parameters, and return types. Routes function calls from Z.AI models to actual implementations (either local MCP tools or external APIs), handles parameter validation, execution, and result marshaling back to model for further reasoning.

Exposes web search capability through MCP tool interface, accepting search queries and returning ranked results with semantic relevance scoring. Implements query expansion, result deduplication, and optional snippet extraction with source attribution. Integrates with Z.AI language models to enable fact-grounded generation and real-time information access.

Implements streaming execution of tool calls through MCP protocol, allowing long-running tools to return results incrementally rather than blocking until completion. Uses server-sent events or chunked transfer encoding to deliver partial results, enabling real-time feedback and progressive refinement. Handles backpressure, error recovery, and result aggregation for tools that produce large outputs (e.g., document processing, video generation).

Implements intelligent routing between Z.AI's two API endpoints: general-purpose endpoint (https://api.z.ai/api/paas/v4) for standard use cases and coding-specific endpoint (https://api.z.ai/api/coding/paas/v4) for code generation, debugging, and analysis tasks. Routes requests based on tool type, model capability, or explicit endpoint selection. Handles endpoint-specific authentication, rate limiting, and error responses.