together-ai-mcp

A Node.js Model Context Protocol (MCP) server that exposes Together AI's inference endpoints — chat completions, image generation, vision, and embeddings — as tools callable from Claude Desktop, Cursor, VS Code, and any other MCP-compatible client.

Why this exists

I created this MCP due to several issues I was having accessing models through Together AI.

1. Reasoning model silent empty responses

Together AI's largest reasoning models (GLM-5, Qwen3.5-397B, MiniMax M2.5, Kimi K2.5) write their chain-of-thought into non-standard response fields, and they exhaust the OpenAI SDK's default token budget before producing a final answer.

Two problems compound each other:

Token budget exhaustion. The OpenAI SDK sets a default max_tokens of 2048. For reasoning models, this budget is consumed entirely by the thinking phase — message.content is never populated. You get charged for tokens, no error is raised, and the response is silently empty.

Fragmented response fields. Different model families on Together AI write their output to different fields:

Any code that only reads message.content — or even message.content \|\| message.reasoning — silently returns an empty string for some models.

// Broken — misses reasoning_content (DeepSeek format):
const text = message.content || message.reasoning || '';

// Fixed — covers all Together AI reasoning model formats:
const text = message.content || message.reasoning_content || message.reasoning || '';

The default max_tokens is raised to 8192 to give reasoning models enough budget to complete their chain of thought before producing a final answer.

2. Vision model failures

Using the OpenAI SDK's chat.completions.create() for vision requests fails silently against Together AI's vision API. Together AI requires stream: false to be set explicitly; the SDK may not send it. When it does fail, the SDK error contains no response body, making the root cause invisible.

// Broken — SDK may omit stream:false; errors are opaque:
const response = await openai.chat.completions.create({ model, messages });

// Fixed — raw fetch, explicit stream:false, full error body in exception:
const response = await fetch('https://api.together.xyz/v1/chat/completions', {
  method: 'POST',
  headers: { Authorization: `Bearer ${apiKey}`, 'Content-Type': 'application/json' },
  body: JSON.stringify({ model, messages, max_tokens, stream: false }),
});
if (!response.ok) {
  const body = await response.text();
  throw new Error(`Vision API error ${response.status}: ${body.slice(0, 200)}`);
}

Related MCP server: Vox MCP

Features

• Chat completions — any Together AI text or reasoning model, with full prompt and multi-turn message support

• Reasoning model support — correctly handles GLM-5, Qwen3.5-397B, MiniMax M2.5, Kimi K2.5 (see above)

• Image generation — FLUX.1-dev, FLUX.1-schnell, Stable Diffusion XL; images saved to disk

• Vision — analyse images via Llama 3.2 Vision or Qwen 2.5 VL

• Embeddings — generate vectors for RAG/retrieval pipelines via BGE and Snowflake Arctic models

Installation

Prerequisites

• Node.js 18+

• A Together AI API key — get one at api.together.ai

Setup

git clone https://github.com/your-username/together-ai-mcp
cd together-ai-mcp
npm install
cp .env.example .env
# Edit .env and add your TOGETHER_API_KEY

Add to Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "together-ai": {
      "command": "node",
      "args": ["/absolute/path/to/together-ai-mcp/index.js"],
      "env": {
        "TOGETHER_API_KEY": "your_api_key_here",
        "IMAGE_OUTPUT_DIR": "/path/to/save/images"
      }
    }
  }
}

See examples/claude-config.md for Cursor and VS Code configuration.

Tools

together_chat

Call any Together AI chat or reasoning model.

together_generate_image

Generate images using FLUX or SDXL models.

Images are saved as PNG files to IMAGE_OUTPUT_DIR.

Note: Image generation uses a direct fetch call rather than the OpenAI SDK's images.generate() because the SDK strips custom parameters like steps when calling Together AI's endpoint.

together_vision

Analyse an image using a vision model.

together_embed

Generate text embeddings for RAG and retrieval pipelines.

Models

The server works with any model available on Together AI's serverless API — just pass its model ID. No configuration changes are needed.

The tables below list the models I personally use. They are provided as a reference, not as a hard limit.

Finding model IDs

Browse all available models at api.together.ai/models. Each model's page shows its exact ID string. Pass that ID as the model parameter to any tool:

{
  "tool": "together_chat",
  "params": {
    "model": "any-model-id-from-together-ai",
    "prompt": "Hello"
  }
}

The only constraint is that image generation models must be called via together_generate_image, vision models via together_vision, and embedding models via together_embed — you cannot call an image model through together_chat.

Dedicated endpoints: Some models (e.g. meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8) require a dedicated endpoint rather than the serverless API. Calling these via this server will return a 400 error from Together AI.

Models I use

Chat / Reasoning

Image generation

Vision

Embeddings

Running tests

npm test

The test suite uses Node.js's built-in test runner and mocks all external dependencies — no API key required to run tests.

Project structure

together-ai-mcp/
├── index.js              # MCP server and handler logic
├── package.json
├── .env.example
├── test/
│   └── index.test.js     # Full test suite (node:test, no external framework)
└── examples/
    ├── chat.md           # Example prompts for each tool and model
    └── claude-config.md  # Configuration for Claude Desktop, Cursor, VS Code

Dependencies

• @modelcontextprotocol/sdk — MCP server framework

• openai — OpenAI-compatible client used with Together AI's baseURL

License

MIT