Oh My Pi (omp) is a terminal coding agent forked from Pi and rewritten as a coding-first harness. It adds IDE-level tooling to the agent loop: language servers on every write, a debugger the agent drives, persistent Python and JavaScript execution, subagents, and in-process search and shell utilities. The same binary runs on macOS, Linux, and Windows.
Edits use hashline patches: the model points at content-hash anchors instead of retyping the lines it changes, and a patch against a stale file is rejected before it is written. Reported results for the edit format include Grok Code Fast 1 rising from 6.7% to 68.3% and Grok 4 Fast using 61% fewer output tokens.
A Rust core of roughly 80,000 lines, shipped as a native addon, runs search, the bash engine, AST matching, syntax highlighting, desktop control, and token counting in-process, so the agent does not shell out to ripgrep, find, or coreutils. A vendored bash fork carries 58 command-line utilities as builtins.
Nine model roles route work by intent (default, smol, slow, plan, commit, vision, task, advisor, and tiny), with per-role fallback chains, path-scoped model sets, and rotating API keys across more than 60 providers. The same engine runs as an interactive TUI, a one-shot prompt, a Node SDK, an NDJSON RPC mode over stdio, and an Agent Client Protocol (ACP) agent for editors such as Zed.
On first run omp reads the rules, skills, and MCP servers other coding agents already wrote to disk (Cursor, Cline, Codex, Copilot, Windsurf, Gemini, and Claude formats), so existing project configuration carries over. Extensions are TypeScript modules with the same tool API and slash-command registry as the built-ins.
Features
- Hashline edits: patches anchor on content hashes, and edits against a stale file are rejected before anything is written
- LSP on every write: diagnostics, navigation, symbols, code actions, and renames that update re-exports and aliased imports
- Debugger control: drives Debug Adapter Protocol sessions (lldb, dlv, debugpy) with breakpoints, stepping, threads, and variables
- Code execution: persistent Python and Bun kernels whose cells can call back into the agent's own read, search, and task tools
- Subagents: the task tool fans work out into isolated worktrees and returns schema-validated results, with an Agent Hub to watch and steer each worker
- Advisor model: a second model reads every turn and injects notes, concerns, or blockers that the main agent responds to
- Stream rules: a regex match aborts the stream mid-token, injects the rule as a reminder, and retries from the same point
- Code review and commits: /review ranks issues P0 to P3 with a verdict, and omp commit splits the working tree into dependency-ordered atomic commits
- Web search: one tool chains 23 search backends and converts GitHub, package registries, arXiv, forums, and docs into structured Markdown
- Unified paths: internal schemes such as pr://, issue://, agent://, skill://, ssh://, and conflict:// work inside read, grep, and write
- Browser and desktop: Puppeteer tabs over Chromium or your own Chrome, plus host desktop control with screenshots, native input, and the accessibility tree
- Memory and collaboration: retain, recall, and learn tools over a local, Hindsight, or Mnemopi backend, and /collab shares a live session by link
