A raw LLM in a browser tab is just a text predictor—what turns models like DeepSeek, Gemini, ChatGPT, or Claude into real autonomous agents is the agent harness around them. I built Web MCP Bridge to wrap their free web interfaces in a complete local Model Context Protocol (MCP) harness using a Manifest V3 extension and a lightweight Python WebSocket bridge (127.0.0.1). By plugging in local stdio MCP servers—including REA for reverse engineering, Craft Anything for 3D modeling, FFmpeg/Remotion for video editing, or full filesystem and terminal access—you can code, edit videos, model 3D assets, and automate complex workflows with zero API keys, zero subscriptions, and zero local GPUs required.
In v1.8.0, the bridge scales to hundreds of tools without freezing browser tabs by replacing massive 100k-character schema dumps with a compact ~2k-character paginated catalog (list_tools) that models query on demand. It also introduces a 3-layer self-healing JSON parser to fix broken model outputs (such as unescaped Windows paths or trailing commas), a redesigned bilingual UI (EN/ES), and dual licensing under MIT and Apache 2.0.
Everything runs 100% locally on your machine: the bridge rejects non-extension origins, isolates tool outputs to prevent prompt injection, and defaults to "Ask every time" so nothing runs without your explicit approval. Grab the code and quick-start guide on GitHub: AstraxOf/web-mcp-bridge
Top comments (0)