BifrostMCP
VS Code language intelligence for AI tools via MCP
205 GitHub stars, 1,600+ extension installs without marketing
Video Demo
AI coding tools have a fundamental information problem. They read text files, infer structure from symbols and naming, and make educated guesses about what calls what and where things are defined. VS Code already knows all of this exactly. Every time you hit go-to-definition or find-all-references, the language server answers in milliseconds with precise, type-checked answers. That information just was not accessible to AI tools.
BifrostMCP is a VS Code extension that acts as an MCP server, bridging that gap. It exposes language server features directly to any AI tool that supports the Model Context Protocol: go-to-definition, find-all-references, get-hover-info, rename-symbol, get-call-hierarchy, find-implementations, get-completions, and about a dozen others. Instead of an AI guessing where a function is defined or what implements an interface, it can ask VS Code and get the exact answer.
Why this did not exist yet
The VS Code Extension API and the Language Server Protocol are genuinely complex. Most AI tool integrations take the simpler path of reading files directly. Getting the extension to act as an MCP server required bridging three different protocol layers: the VS Code Extension API, LSP, and MCP. The extension needs to stay alive in the background, handle requests from external tools over the MCP transport, translate those into VS Code API calls, and return structured results. Getting this working reliably across different language servers (TypeScript, C#, Python, etc.) required working through a lot of edge cases in how different servers respond to the same LSP requests.
Adoption without marketing
I posted it in a few relevant communities and that was it. 205 GitHub stars and 1,600+ VS Code extension installs came from developers discovering it organically. Developers using Claude, Cursor, and other MCP-compatible tools have picked it up because it solves a real problem they hit immediately when trying to use AI tools on large codebases. When a tool genuinely fills a gap, it finds its users.