Oh-My-Pi: An LSP and DAP-Aware AI Agent Harness

Better Stack2026-05-30go watch the original →

the gist

Oh-My-Pi replaces standard text-based AI CLI wrappers with an architecture that integrates directly with IDE protocols and uses content-hash anchoring to reduce token usage by 61%.

Architectural Improvements for AI Agents

Oh-My-Pi shifts the paradigm of terminal-based AI agents by treating projects as active application runtimes rather than static collections of text files. This approach allows the agent to perform structural refactoring and debugging with the same context as a full-featured IDE.

Key Technical Features

Native LSP Integration: The agent utilizes the Language Server Protocol to perform workspace-wide refactors, such as renaming modules or updating imports across multiple files, ensuring consistency in barrel files and alias imports.
Debugger Adapter Protocol (DAP) Support: It integrates with debuggers like DLV or debugpy to attach to running processes, hit breakpoints, and inspect live memory states or stack frames.
Content-Hash Anchoring: Instead of sending full file diffs, the system targets specific lines using content hashes. This method prevents syntax errors and reduces LLM token consumption by up to 61% compared to standard text-replacement methods.
Model Agnostic Architecture: Users can configure different models for specific tasks, such as using dedicated vision models for UI analysis or specialized models for architectural design, while maintaining compatibility with existing Claude Code plugins.
Integrated Browser Tooling: The agent includes a headless Chrome instance for web data retrieval, avoiding the limitations of simple curl or fetch requests.

Before / After

Token Efficiency: By using content-hash anchoring instead of full-string diffs, the system achieves a 61% reduction in LLM token usage during file edits.