convez.ai
MCP ServerIntegração & APIsIntermediário
xiaobaotalks

Servidor MCP OpenAI-Compatível

xiaobaotalks/openai-compatible-mcp

Bridge stdio MCP para qualquer API OpenAI-compatível

Ver no GitHub
InicialPluginScore · de 0 a 100

Servidor MCP que funciona como ponte para qualquer API de chat compatível com OpenAI (DeepSeek por padrão). Expõe ferramentas de chat completion e listagem de modelos com suporte a múltiplos provedores como OpenAI, Azure, local llama.cpp e mais. Ideal para integrar diferentes LLMs em clientes MCP como Claude Desktop e Cursor.

Instalação

$npx convez-ai xiaobaotalks/openai-compatible-mcp

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 3 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Integrar DeepSeek, OpenAI ou outros LLMs compatíveis em Claude Desktop
  • Usar modelos locais (llama.cpp) como backend de chat em aplicações MCP
  • Abstrair múltiplos provedores de API com uma única interface

Veredito

Excelente ferramenta minimalista (400 linhas) que resolve um problema real: unificar acesso a APIs OpenAI-compatíveis sem dependências externas. Recomendado para quem quer máxima flexibilidade na escolha de modelos em ecossistema MCP.

README

A tradução para o português deste README ainda está na fila.

openai-compatible-mcp

A Model Context Protocol (MCP) server that bridges to any OpenAI-compatible chat API — DeepSeek, OpenAI, Azure OpenAI, OpenRouter, Together, Groq, local llama.cpp, and friends.

Default provider: DeepSeek. Switch to any OpenAI-compatible endpoint with a single environment variable.

The server runs over stdio (the standard MCP transport) and exposes two tools that any MCP-compatible client can call:

  • chat — send a chat completion, get the assistant's reply
  • list_models — show the configured default model and friendly aliases

Features

  • Zero third-party dependencies (Python 3.9+ stdlib only)
  • Works with any OpenAI-compatible chat API
  • Friendly model aliases (deepseek-v4-flash → deepseek-chat, r1 → deepseek-reasoner, etc.)
  • Optional reasoning content extraction (DeepSeek-R1 style)
  • Built-in JSON-RPC 2.0 over stdio MCP transport
  • Tiny: ~400 lines of code

Quick start

1. Install

From source:

git clone https://github.com/xiaobaotalks/openai-compatible-mcp.git
cd openai-compatible-mcp
pip install -e .

Or run directly without installing:

PYTHONPATH=src python -m openai_compatible_mcp

2. Configure

Set your API key and (optionally) the base URL:

# DeepSeek (default)
export DEEPSEEK_API_KEY="sk-..."

# OpenAI
export OPENAI_API_KEY="sk-..."
export OPENAI_COMPATIBLE_MCP_BASE_URL="https://api.openai.com"

# Azure OpenAI
export OPENAI_COMPATIBLE_MCP_API_KEY="..."
export OPENAI_COMPATIBLE_MCP_BASE_URL="https://YOUR-RESOURCE.openai.azure.com/openai/deployments/YOUR-DEPLOYMENT"

# Local llama.cpp server
export OPENAI_COMPATIBLE_MCP_BASE_URL="http://127.0.0.1:8080"
export OPENAI_COMPATIBLE_MCP_DEFAULT_MODEL="llama-3.1-8b"

# Custom: anything speaking the OpenAI /v1/chat/completions protocol
export OPENAI_COMPATIBLE_MCP_API_KEY="..."
export OPENAI_COMPATIBLE_MCP_BASE_URL="https://my.example.com"
export OPENAI_COMPATIBLE_MCP_DEFAULT_MODEL="my-model"

Environment variable lookup order (first one set wins):

SettingVariables tried (in order)
API keyOPENAI_COMPATIBLE_MCP_API_KEY → DEEPSEEK_API_KEY → OPENAI_API_KEY
Base URLOPENAI_COMPATIBLE_MCP_BASE_URL → DEEPSEEK_API_BASE → OPENAI_BASE_URL
Default modelOPENAI_COMPATIBLE_MCP_DEFAULT_MODEL → DEEPSEEK_DEFAULT_MODEL

3. Wire it up to an MCP client

Claude Desktop

~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "openai-compatible": {
      "command": "python",
      "args": ["-m", "openai_compatible_mcp"],
      "env": {
        "DEEPSEEK_API_KEY": "sk-..."
      }
    }
  }
}

See examples/claude_desktop_config.json for a full sample.

Cursor

~/.cursor/mcp.json:

{
  "mcpServers": {
    "openai-compatible": {
      "command": "python",
      "args": ["-m", "openai_compatible_mcp"],
      "env": {
        "DEEPSEEK_API_KEY": "sk-..."
      }
    }
  }
}

Claude Code / other stdio clients

Same shape: launch python -m openai_compatible_mcp as a subprocess and read/write JSON-RPC over its stdin/stdout.

Tools

chat

Send a chat completion to the configured provider.

FieldTypeRequiredDescription
messagesarrayyesList of {role, content} messages (oldest first).
modelstringnoModel name or alias (e.g. deepseek-v4-flash, o1-mini).
temperaturenumberno0-2, lower = more deterministic.
max_tokensintegernoMaximum tokens to generate.
top_pnumbernoNucleus sampling cutoff.
stopstring | arraynoStop sequence(s).
systemstringnoSystem prompt (prepended to the conversation).
include_reasoningbooleannoWrap reasoning content in <think>...</think>.

Example (from an MCP client):

{
  "method": "tools/call",
  "params": {
    "name": "chat",
    "arguments": {
      "messages": [
        {"role": "user", "content": "Write a haiku about Python."}
      ],
      "model": "deepseek-v4-flash",
      "temperature": 0.7
    }
  }
}

The tool returns text content with the assistant's reply plus a usage footer:

Whitespace glides through the code,
Indents bloom like spring leaves—
Bugs hide, then they are gone.

---
model: deepseek-chat | prompt_tokens: 12 | completion_tokens: 28 | total_tokens: 40

list_models

Returns the default model and the alias table.

Built-in model aliases

You can edit src/openai_compatible_mcp/client.py to add or remove aliases for your provider. Defaults:

AliasResolves to
deepseek-v4-flashdeepseek-chat
deepseek-v4-prodeepseek-chat
deepseek-v3deepseek-chat
deepseek-chatdeepseek-chat
deepseek-reasonerdeepseek-reasoner
deepseek-r1deepseek-reasoner
deepseek-coderdeepseek-chat
gpt-4ogpt-4o
gpt-4o-minigpt-4o-mini
o1o1
o1-minio1-mini
o3-minio3-mini

Any model name not in the alias table is passed through to the provider unchanged, so the latest models work the moment a provider releases them.

Smoke test

Run the server and send a few raw JSON-RPC requests over stdio:

echo '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"smoke","version":"0"}}}' | python -m openai_compatible_mcp

Or use the included test client:

PYTHONPATH=src python tests/smoke_test.py

Development

# Create a venv
python -m venv .venv
.venv\Scripts\activate       # Windows
source .venv/bin/activate    # macOS / Linux

# Editable install
pip install -e ".[dev]"

# Run tests
PYTHONPATH=src python tests/smoke_test.py

License

MIT — see LICENSE.

Ver o original no GitHub

Bom para

Relacionados