agent-lsp

agent-lsp

MCP server for language intelligence. Give AI agents structured access to language servers: go-to-definition, find-all-references, rename, diagnostics, completions, call hierarchy, and speculative execution.

66 tools. 23 agent workflows. 30 CI-verified languages. Single Go binary.

Why

AI coding agents navigate code with grep and file reads. This wastes 5-34x more tokens than necessary and produces 92-99% false positives on symbol lookups. Language servers already have the answers; agent-lsp makes them accessible via MCP.

Install

pip install agent-lsp

Or via other channels:

brew install blackwell-systems/tap/agent-lsp    # macOS/Linux
npm install -g @blackwell-systems/agent-lsp     # npm
curl -fsSL https://raw.githubusercontent.com/blackwell-systems/agent-lsp/main/install.sh | sh

Usage

# Start with auto-detection
agent-lsp

# Explicit language server
agent-lsp go:gopls typescript:typescript-language-server,--stdio

# HTTP mode
agent-lsp --http --port 8080 go:gopls

Then configure your AI tool's MCP settings to point at agent-lsp.

What it does

• 66 tools covering navigation, analysis, refactoring, diagnostics, formatting, speculative execution, composite exploration, safe editing, build, test, and more
• 23 agent workflows (skills) that encode correct multi-step operations: rename safely, analyze blast radius, simulate edits before applying
• 30 CI-verified languages: Go, Python, TypeScript, Rust, Java, C, C++, C#, Ruby, PHP, Kotlin, Swift, Scala, Zig, Lua, Elixir, Gleam, and more
• Speculative execution: preview edits in memory, see what breaks before touching disk
• Phase enforcement: blocks out-of-order operations so agents follow correct workflows

Token savings

Measured across 5 codebases (Go, TypeScript, Python):

Full experiment: agent-lsp.com/token-savings

Works with

Claude Code, Cursor, Windsurf, GitHub Copilot, and any MCP-compatible client.

Links

• GitHub
• Documentation
• Token savings experiment
• Changelog