agent-lsp

agent-lsp is a Model Context Protocol (MCP) server designed to provide language intelligence to AI agents by bridging the Language Server Protocol (LSP) with MCP-based workflows. The tool addresses the challenge of AI agents making incorrect code modifications due to limited context, such as not knowing who calls a function or what breaks if a change is made. By connecting AI agents with real language servers, agent-lsp enables them to access code intelligence features like completions, diagnostics, go-to-definition, and more, across a wide range of programming languages.

The platform operates as a stateful runtime that maintains a warm index of the workspace, avoiding the inefficiency of cold-starting on every request. It supports 66 tools and 30 languages, with compatibility verified through continuous integration tests using real language servers and fixture codebases. Supported languages include Go, Python, TypeScript, Rust, Java, C, C++, C, Ruby, PHP, Kotlin, Swift, Scala, Zig, Lua, Elixir, Gleam, Clojure, Dart, Terraform, Nix, Prisma, SQL, MongoDB, among others. agent-lsp also provides 24 agent workflows, referred to as skills, which are multi-step procedures discoverable via MCP prompts or lists.

A notable feature is speculative execution, allowing agents to simulate code changes in memory before committing them to disk. This includes previewing the diagnostic impact of edits and evaluating sequences of dependent changes to identify where errors may be introduced. The tool also enforces workflow phase ordering at runtime through phase enforcement, ensuring that agents follow the correct sequence of operations and providing recovery guidance if violations occur. This approach helps maintain code integrity during automated agent-driven workflows.

agent-lsp is delivered as a single Go binary and offers multiple installation channels. It supports persistent connections, maintains a warm index across files and projects, and features HTTP+SSE transport with bearer token authentication. The tool integrates with various AI tools and MCP clients via stdio or HTTP, and can be configured for use with platforms like Claude Code, Continue, Cline, and Cursor. The architecture is designed for ease of integration and continuous, reliable operation in agent-driven code intelligence scenarios.