# oh-my-pi
oh-my-pi is an open-source terminal coding agent maintained by Can Bölük (Stencil Labs). The CLI is called `omp`. It's MIT licensed, written in TypeScript with a native [[Rust]] core of about 80k lines, and runs on [[Bun]]. The site describes it as "A coding agent with the IDE wired in."
It started as a fork of [[Pi Mono]] by [[Mario Zechner]] ("Originally built on Mario Zechner's wonderful Pi"), then got rewritten as a coding-first, batteries-included tool. That's the opposite of Pi's rule ("If I don't need it, it won't be built"), and Pi still ships without subagents or a plan mode. As of October 2026, oh-my-pi comes with 31 built-in tools, 14 LSP operations, 28 DAP operations and more than 60 model providers.
## Editor tooling
- [[Language Server Protocol (LSP)]] runs on every write. Renames go through `workspace/willRenameFiles`, so imports get updated
- A debugger tool over DAP (lldb, delve, debugpy)
- Persistent Python and Bun/JS kernels that can call back into the agent's tools
- `ast_grep` and `ast_edit` over 50+ [[Tree-sitter]] grammars; `ast_edit` shows a preview before applying
- Ripgrep-style search, a persistent bash (brush) and coreutils all run in-process, with no fork/exec
## Agent features
Edits target content-hash anchors ("hashline edits"). A stale anchor gets rejected instead of corrupting the file. The README claims this saves output tokens and retry loops; those numbers come from the project itself.
Time-traveling stream rules (TTSR) watch the output with regexes. On a match, the stream is aborted mid-token, a reminder gets injected and the model retries. Injections survive compaction.
Other features:
- [[AI Subagents]] through a `task` tool, optionally in a [[Git Worktree]], with schema-validated results
- An advisor role: a second model reviews each turn and adds notes
- `/review` runs parallel reviewer subagents that rank issues from P0 to P3, and `/collab` shares a live session (read-write or view-only)
- Cross-session memory with pluggable backends (local, Hindsight, Mnemopi), web search across 23 providers, browser and desktop control
- It reads rules, [[AI Agent Skills]] and [[Model Context Protocol (MCP)]] servers from other tools' folders (`.claude`, `.cursor`, `.codex`)
- [[Model routing]] per role (default, smol, slow, plan, commit, vision…), fallback chains and round-robin credentials
- TypeScript extensions that use the same APIs as the built-in tools
## Usage and install
You can use it through the interactive TUI, one-shot prompts (`-p`), a Node SDK, JSON-over-stdio RPC, or [[Agent Client Protocol (ACP)]] in editors like Zed. It runs on macOS and Linux: `curl -fsSL https://omp.sh/install | sh`. Homebrew, bun, Nix, PowerShell and mise are supported as well.
Project-level settings, extensions, hooks and MCP configuration load by default. So opening an untrusted repository with `omp` means running whatever that repository configures.
## Why it matters
To me, Pi and oh-my-pi sit at the two ends of [[Harness Engineering]]. Pi gives you a small, observable loop and you build what you need on top. oh-my-pi bets on a complete [[AI Agent Harness]] where the IDE side (LSP, debugger, AST edits) is built in, with "Every tool tuned against real sessions". Its own pitch is "A harness worth keeping is one you don't outgrow". The trade-off is a much larger surface to trust.
## References
- Website: https://omp.sh
- Documentation: https://omp.sh/docs
- Source code: https://github.com/can1357/oh-my-pi
## Related
- [[Pi Mono]]
- [[Mario Zechner]]
- [[AI Agents]]
- [[AI Agent Harness]]
- [[Harness Engineering]]
- [[AI Subagents]]
- [[AI Agent Skills]]
- [[Claude Code]]
- [[OpenCode]]
- [[Codex CLI]]
- [[Gemini CLI]]
- [[Language Server Protocol (LSP)]]
- [[Model Context Protocol (MCP)]]
- [[Agent Client Protocol (ACP)]]
- [[Tree-sitter]]
- [[Git Worktree]]
- [[Bun]]
- [[Rust]]