Oh My Pi

• Tracking down a segfault, hang, or wrong value by having the agent attach a debugger and inspect the live process instead of adding print statements
• Running a cross-package refactor or rename that relies on the language server to update imports and re-exports
• Splitting an audit or migration across parallel subagents in isolated workspaces, then integrating and verifying the results in the main session
• Pairing on a live agent session with a teammate, or sharing a view-only link for a review or demo
• Reviewing a branch or uncommitted work with prioritized findings, then splitting the change into atomic commits
• Moving from Claude Code, Cursor, or Codex without rewriting existing rules, skills, and MCP server configs

omp is what you get when someone takes Pi's terminal workflow and decides the agent should already have everything an IDE has. Language servers, a real debugger, structural edits, subagents with isolated workspaces, browser control, and shared sessions are all built in. Hash-anchored edits target the whitespace and string-not-found retries that the project says slow other agents down. The cost is surface area and defaults. There is a lot to learn, releases land several times a week, approvals are off until you turn them on, and repository config loads automatically. It suits developers who will set up approvals and project trust before relying on it and who want one tool to grow into. Anyone who chose Pi for its small core should stay with Pi.

About Oh My Pi

Oh My Pi started as a fork of Pi (pi-mono, the open-source toolkit by Mario Zechner that is now maintained by Earendil) and was rewritten as a coding-first agent. The two now take opposite positions. Pi keeps a small core: MCP is now built in, but sub-agents, plan mode, and permission prompts are left to extensions. omp puts all of that in the box. You run `omp` in a repository, describe the outcome, and the agent works through a large built-in tool set. Edits use hashline patches that point at content-hash anchors instead of retyping lines, and a stale anchor makes the patch fail instead of corrupting the file. Renames and diagnostics go through the project's language servers, and the agent can launch or attach a debugger through the Debug Adapter Protocol (for example lldb, dlv for Go, or debugpy for Python) to set breakpoints and read live variables. Persistent Python and JavaScript cells can call back into the agent's own tools. Structural search and rewrites use ast-grep. A `browser` tool drives headless Chromium or your own Chrome tabs through a relay extension, an optional `computer` tool controls the desktop, and `web_search` chains many search backends. The `task` tool fans work out to subagents, optionally in isolated workspaces, and returns schema-validated results. You can watch and steer each worker in Agent Hub. An optional advisor model can review every turn, /review runs prioritized code review with reviewer subagents, and `omp commit` splits a working tree into atomic commits. /collab shares a live session through an end-to-end encrypted relay, either read-write or view-only, with guests joining from another terminal or a browser. omp works with dozens of providers through API keys, OAuth sign-in (including Claude Pro/Max and ChatGPT subscriptions), coding plans such as GitHub Copilot, and local engines such as Ollama, LM Studio, and llama.cpp. You assign models to roles such as default, smol, slow, and plan. On first run it picks up existing rules, skills, and MCP servers from Claude Code, Cursor, Codex, Cline, Windsurf, Gemini, Copilot, and VS Code config folders. Besides the TUI it runs as a one-shot print command, an NDJSON RPC server, a Node SDK, and an Agent Client Protocol agent for editors such as Zed. The tradeoffs are breadth and defaults. The surface is large and changes several times a week. The built-in approval mode is yolo, so tool calls run without prompts until you configure approvals (ACP sessions keep the editor's permission prompt for shell commands and edits). Opening a repository also loads its settings, hooks, extensions, and MCP configuration by design.