🧠 Cerebras Multi-Model MCP Server

Use multiple Cerebras models from Claude Desktop & Claude Code — with automatic model selection.

---

The Problem

The official Cerebras MCP package only supports one model per session — you pick a model via an environment variable, and you're stuck with it until you restart. Want to use the fast 8B model for boilerplate and the 357B model for complex reasoning? You'd need two separate MCP server configs.

The Solution

cerebras-multi-mcp exposes 5 tools — one for each Cerebras model plus an auto-selector — so you (or Claude) can pick the right model per task, in the same session, with zero restarts.

┌─────────────────────────────────────────────────┐
│              Claude Desktop / Code               │
├─────────────────────────────────────────────────┤
│                                                  │
│  cerebras_quick     → llama3.1-8b      (8B)     │
│  cerebras_complex   → gpt-oss-120b     (120B)   │
│  cerebras_reasoning → zai-glm-4.7      (357B)   │
│  cerebras_instruct  → qwen-3-235b      (235B)   │
│  cerebras_auto      → picks the best one        │
│                                                  │
├─────────────────────────────────────────────────┤
│         Cerebras API  ←→  OpenRouter Fallback    │
└─────────────────────────────────────────────────┘

---

Models

Tool	Model	Params	Best For
cerebras_quick	llama3.1-8b	8B	Simple edits, boilerplate, single functions. Fastest.
cerebras_complex	gpt-oss-120b	120B	Multi-file features, CRUD APIs, complex components.
cerebras_reasoning	zai-glm-4.7	357B	Algorithms, architecture, advanced logic, deep reasoning.
cerebras_instruct	qwen-3-235b	235B	Precise instructions, documentation, typed interfaces, specs.
cerebras_auto	auto-selected	—	Analyzes your prompt and picks the best model automatically.

Auto-Selection Logic

cerebras_auto analyzes your prompt keywords and complexity:

• Reasoning keywords (algorithm, optimize, recursive, big-o…) → 357B
• Instruct keywords (document, jsdoc, schema, openapi…) → 235B
• Complex keywords (crud, rest api, multi-file, database…) → 120B
• Everything else or short prompts → 8B (fastest)

---

Installation

Prerequisites

• Node.js 18+
• A Cerebras API key (free tier available)
• (Optional) An OpenRouter API key for fallback

Setup

git clone https://github.com/khansabassem/cerebras-multi-mcp.git
cd cerebras-multi-mcp
npm install

---

Configuration

Claude Desktop

Edit your Claude Desktop config file:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

Add the cerebras-multi entry:

{
  "mcpServers": {
    "cerebras-multi": {
      "command": "node",
      "args": ["<path-to>/cerebras-multi-mcp/src/index.js"],
      "env": {
        "CEREBRAS_API_KEY": "your-cerebras-api-key",
        "OPENROUTER_API_KEY": "your-openrouter-api-key"
      }
    }
  }
}

Restart Claude Desktop to load the new server.

Claude Code

claude mcp add cerebras-multi \
  -e CEREBRAS_API_KEY=your-cerebras-api-key \
  -e OPENROUTER_API_KEY=your-openrouter-api-key \
  -- node /path/to/cerebras-multi-mcp/src/index.js

---

Usage

Once configured, you'll see 5 new tools in Claude. Each tool accepts:

Parameter	Required	Description
file_path	Yes	Absolute path to the file to create or modify
prompt	Yes	Detailed code generation instructions
context_files	No	Array of file paths to read as context
temperature	No	Sampling temperature (default: 0.1)
max_tokens	No	Maximum tokens in the response

Examples

Quick boilerplate with the 8B model:

Tool: cerebras_quick
file_path: /project/src/server.js
prompt: Create an Express server with health check endpoint on port 3000

Complex feature with the 120B model:

Tool: cerebras_complex
file_path: /project/src/auth/middleware.ts
prompt: Create JWT authentication middleware with refresh token rotation
context_files: ["/project/src/types/auth.ts", "/project/src/config/env.ts"]

Algorithm design with the 357B model:

Tool: cerebras_reasoning
file_path: /project/src/utils/graph.ts
prompt: Implement Dijkstra's shortest path with a priority queue, supporting weighted directed graphs

Documentation with the 235B model:

Tool: cerebras_instruct
file_path: /project/src/types/api.ts
prompt: Generate TypeScript interfaces for a REST API with OpenAPI-compatible JSDoc annotations

Let the server decide:

Tool: cerebras_auto
file_path: /project/src/cache.ts
prompt: Build an LRU cache with O(1) get and put using a doubly linked list

---

Features

• Per-call model selection — no restarts, no env var juggling
• Auto-select mode — keyword analysis picks the right model for you
• OpenRouter fallback — if Cerebras is unavailable, requests fall through to OpenRouter
• Smart file handling — reads existing files for context when editing, creates directories as needed
• Diff summaries — shows additions/removals when updating existing files
• Code cleaning — strips markdown fences from model output automatically
• Context files — pass related files for cross-file awareness

---

Architecture

src/index.js          — Single-file MCP server (~350 lines)
├── Config            — Model definitions, keyword lists, language detection
├── File helpers      — Safe read/write with path resolution
├── HTTP layer        — Cerebras API + OpenRouter fallback
├── Auto-selector     — Keyword-based model routing
├── Tool handler      — Unified handler for all 5 tools
└── MCP server        — ListTools + CallTool with schema factory

Built with @modelcontextprotocol/sdk using stdio transport.

---

Why Cerebras?

Cerebras inference runs on purpose-built wafer-scale hardware, delivering up to 20x faster inference than traditional GPU setups. Combined with MCP, you get near-instant code generation directly inside Claude.

---

Author

Bassem EL KHANSAA — @ask.bassem

---

License

MIT

---

Contributing

Issues and PRs welcome. If you add a new model, just extend the MODELS object and add a tool entry in the ListToolsRequestSchema handler.