AI coding agents today use grep/ripgrep for code search. This fails on real codebases: can't find references across modules, can't preview refactorings, can't understand code structure. LSP (Language Server Protocol) solves this for human IDEs. But existing agent integrations (Serena, cclsp, Claude Code LSP, etc.) just expose raw LSP endpoints—agents need 10+ round trips to get usable answers.
We built lsp-skill for this: https://lsp-client.github.io/
Three-layer design:
- lsp-client: Robust LSP client (handles server lifecycle, protocol edge cases, multi-workspace)
- LSAP (Language Server Agent Protocol): Agent-optimized protocol (composes LSP primitives into high-level operations)
- Skill layer: Integrates with Claude Code, Cursor, MCP-compatible tools
try it:
- Docs: https://lsp-client.github.io/
- GitHub: https://github.com/lsp-client/lsp-skill
- Install: Works with Claude Code, Cursor, or any MCP tool