Skip to main content
Glama
Cerebras Multi-Model MCP Server

Cerebras Multi-Model MCP Server

by khansabassem

🧠 Cerebras Multi-Model MCP Server

Use multiple Cerebras models from Claude Desktop & Claude Code β€” with automatic model selection.

MCP Compatible Node.js License: MIT


The Problem

The official Cerebras MCP package only supports one model per session β€” 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.

Related MCP server: Zen MCP Server

The Solution

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

β”Œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”
β”‚              Claude Desktop / Code               β”‚
β”œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€
β”‚                                                  β”‚
β”‚  cerebras_quick     β†’ llama3.1-8b      (8B)     β”‚
β”‚  cerebras_complex   β†’ gpt-oss-120b     (120B)   β”‚
β”‚  cerebras_reasoning β†’ zai-glm-4.7      (357B)   β”‚
β”‚  cerebras_instruct  β†’ qwen-3-235b      (235B)   β”‚
β”‚  cerebras_auto      β†’ picks the best one        β”‚
β”‚                                                  β”‚
β”œβ”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€
β”‚         Cerebras API  ←→  OpenRouter Fallback    β”‚
β””β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”€β”˜

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

β€”

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…) β†’ 357B

  • Instruct keywords (document, jsdoc, schema, openapi…) β†’ 235B

  • Complex keywords (crud, rest api, multi-file, database…) β†’ 120B

  • Everything else or short prompts β†’ 8B (fastest)


Installation

Prerequisites

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 β€” no restarts, no env var juggling

  • Auto-select mode β€” keyword analysis picks the right model for you

  • OpenRouter fallback β€” if Cerebras is unavailable, requests fall through to OpenRouter

  • Smart file handling β€” reads existing files for context when editing, creates directories as needed

  • Diff summaries β€” shows additions/removals when updating existing files

  • Code cleaning β€” strips markdown fences from model output automatically

  • Context files β€” pass related files for cross-file awareness


Architecture

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

GitHub Instagram LinkedIn Website


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.

Available Tools

5 tools
cerebras_autoA

Auto-selects the best Cerebras model based on prompt complexity. Simple tasks use 8B, complex features use 120B, reasoning tasks use 357B, documentation tasks use 235B. Use this when unsure which model fits.

Inputs: prompt (required), file_path (required), max_tokens, temperature, +1 more.

cerebras_complexB

Heavy-duty code generation using Cerebras gpt-oss-120b (120B params). Large model for multi-file features, CRUD APIs, complex components, and demanding code generation tasks.

Inputs: prompt (required), file_path (required), max_tokens, temperature, +1 more.

cerebras_instructB

Instruction-following code generation using Cerebras zai-glm-4.7 (355B params, reasoning_format:hidden). Instruction-tuned model for precise instruction following, documentation-heavy code, typed interfaces, and detailed specs.

Inputs: prompt (required), file_path (required), max_tokens, temperature, +1 more.

cerebras_quickA

Fast code generation using Cerebras gpt-oss-120b (120B params). Fast model for simple edits, boilerplate, single-function generation. Use for trivial tasks where speed matters most.

Inputs: prompt (required), file_path (required), max_tokens, temperature, +1 more.

cerebras_reasoningA

Advanced reasoning code generation using Cerebras zai-glm-4.7 (355B params, reasoning_format:hidden). Most powerful model for algorithms, architecture decisions, advanced logic, and tasks requiring deep reasoning.

Inputs: prompt (required), file_path (required), max_tokens, temperature, +1 more.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A3.8/5.0

    Scored across 5 tools

    Disambiguation5/5

    Each tool targets a distinct use case: auto-selection, heavy code, instruction-following, fast generation, or deep reasoning. Descriptions clearly separate model sizes and tasks, eliminating ambiguity.

    Naming Consistency5/5

    All tool names follow a consistent 'cerebras_<adjective>' pattern using snake_case. The descriptors (auto, complex, instruct, quick, reasoning) clearly indicate the tool's purpose.

    Tool Count5/5

    With 5 tools, the server covers the main model variants without being excessive. The count is well-scoped for a model selection server, providing essential choices without redundancy.

    Completeness4/5

    The tool surface covers auto-selection, heavy tasks, instruction-tuned, fast, and reasoning use cases. One minor gap: no explicit 'chat' or 'vision' tool, but the domain appears focused on code generation, so the set is largely complete.

    Maintenance

    ActivityInactive
    ResponsivenessNo recent community issues

    Related MCP Connectors

    Related MCP Servers

    • -
      license
      Not graded
      quality
      Not graded
      maintenance
      Gives Claude access to multiple AI models (Gemini, OpenAI, OpenRouter, Ollama) for enhanced development capabilities including extended reasoning, collaborative development, code review, and advanced debugging.
      -
    • F
      license
      Not graded
      quality
      Not graded
      maintenance
      Gives Claude Desktop access to multiple AI models (Gemini, OpenAI, OpenRouter, Ollama) for enhanced development capabilities including extended reasoning, collaborative development, code review, debugging, and large context analysis with conversation threading.
      40 npm
      1
      -