ChatLocal

A self-hosted ChatGPT replacement for your locally running models. Connects to any OpenAI-compatible inference endpoint — vLLM, Ollama, LM Studio, llama.cpp server — and gives you a polished chat UI with no data leaving your machine.

Why

ChatGPT and Claude are great, but sometimes you need:

• Full privacy — every token stays on your hardware
• No rate limits — your GPU, your rules
• Bleeding-edge models — run whatever is on Hugging Face today without waiting for an API
• Tool use with local context — connect MCP servers to give the model access to your files, APIs, and services

ChatLocal gives you the ChatGPT-style experience on top of your own inference stack.

Features

• Real-time streaming — tokens appear as they are generated via WebSocket
• Thinking blocks — collapsible reasoning display for models like Qwen3 that emit <think> tokens
• MCP tool integration — connect any MCP server (stdio transport); tools run automatically and appear inline as collapsible cards
• Session management — persistent chat history in SQLite, auto-named after the first exchange
• Full-text search — FTS5-powered search across all messages in the sidebar
• Vision support — attach images via paste, drag-and-drop, or file picker
• Context compression — automatically summarizes old messages when approaching the model's context limit
• Markdown rendering — tables, code blocks with syntax highlighting and one-click copy, links
• Export — download any session as a Markdown file
• Settings — configurable system prompt, temperature, top-p, max tokens

Requirements

• Node.js 20+
• An OpenAI-compatible inference server (vLLM, Ollama, LM Studio, etc.) running locally or on your network

Setup

npm install

Create a .env file:

VLLM_BASE_URL=http://localhost:8000   # your inference server base URL
PORT=3000
DATABASE_PATH=./data/chatlocal.db
TAVILY_API_KEY=                        # optional, for web search via MCP

Start the server:

npm start

Open http://localhost:3000.

MCP Tools

Tool servers are configured in mcp-servers.json. Each entry is a stdio process that ChatLocal spawns on startup. Environment variables are interpolated automatically.

{
  "tavily-remote": {
    "command": "npx",
    "args": ["-y", "mcp-remote", "https://mcp.tavily.com/mcp/?tavilyApiKey=${TAVILY_API_KEY}"]
  }
}

Any MCP server that speaks stdio transport works here. Changes require a server restart.

Connecting to different backends

Backend	VLLM_BASE_URL
vLLM	http://localhost:8000
Ollama	http://localhost:11434
LM Studio	http://localhost:1234
llama.cpp server	http://localhost:8080

The model list is populated automatically from GET /v1/models. If your backend doesn't serve that endpoint, set the model ID manually in the settings.