What is omp (oh-my-pi)?
omp is a terminal coding agent that started as a fork of Mario Zechner's Pi and is now maintained by Can Bölük and Stencil Labs under the MIT licence. Where Pi keeps a small core and leaves features to extensions, omp ships them built in: 31 tools, language-server and debugger access, parallel subagents in isolated worktrees, a second advisor model, web search across 25 backends, browser and desktop control, persistent Python and JavaScript cells, agent-curated memory and shared live sessions.
About 80,000 lines of Rust replace the shell-outs other agents make: ripgrep, glob, an embedded bash 5.3 engine and 58 command-line utilities run in-process, which is also how the same binary runs on Windows without WSL. It runs as a TUI, as a one-shot command, over RPC, as a Node SDK and as an ACP agent inside editors such as Zed.
Is omp (oh-my-pi) worth it?
omp answers the question Pi deliberately leaves open: what if the extensions were already written? It keeps Pi's terminal interface, session tree and TypeScript extension model, then adds what Pi omits on purpose. Renames go through the language server, so imports and re-exports update before a file moves. The agent can attach lldb, dlv or debugpy and step through a crash instead of adding print statements. Subagents fan out into separate git worktrees and return schema-checked results, and a second model in an advisor role can read every turn and flag problems.
The headline engineering claim is the edit format. Instead of retyping lines to replace, the model points at content-hash anchors, and a stale anchor rejects the patch rather than corrupting the file. The author's own benchmark reports Grok Code Fast 1 rising from 6.7% to 68.3% on edits and Grok 4 Fast spending 61% fewer output tokens; these are vendor figures, not independent tests. Provider coverage is wide, at more than 60, including Claude, Codex, Copilot, Cursor and the GLM, Kimi and MiniMax coding plans.
The project moves fast. The repository has 39,533 stars, v19.0.1 shipped on 19 November 2026 a day after v19.0.0, and the npm package had 1,240,980 downloads in the 30 days to 19 November. GitHub counts 4,977 open issues and pull requests, a backlog that signals both popularity and churn.
The catch: the default approval mode is yolo, which runs reads, writes and shell commands without asking. Stricter modes exist (always-ask, ask-and-remember, write) and restricted folders always prompt, but you must turn them on, and there is no sandbox. Opening a repository also loads its settings, hooks, extensions and project MCP servers by design, so an untrusted checkout can run code. Telemetry is limited: automatic tool-issue reports need saved consent and OpenTelemetry export runs only if you configure an endpoint, though a random install ID is sent to some providers and gateways.
Our take: Pi is the kit and omp is the finished workshop. The trade is a larger, faster-changing surface to audit in return for features you would otherwise build yourself.
Where it earns its place
A backend engineer chasing a segfault in a C service
Have the agent find the bad pointer instead of guessing from logs
omp attaches lldb through its debug tool, steps to the failing frame and reads the variables. The agent is free; cost is the model you log in with.
A tech lead splitting a refactor across a TypeScript monorepo
Run several agents in parallel without them overwriting each other
The task tool spawns subagents in separate worktrees and returns structured findings; Alt+A shows each one's progress and cost. Set approvalMode to write or always-ask first.
A developer already paying for a GLM or Kimi coding plan
Use that subscription in a stronger harness than the vendor's own
/login attaches the plan, and fallback chains move a turn to another model when the plan hits its quota. No licence cost.
Before you commit
- Setup
- `curl -fsSL https://omp.sh/install | sh` (the script checks the binary's SHA-256), Homebrew, `bun install -g @oh-my-pi/pi-coding-agent`, Nix or PowerShell; run `omp` and `/login`. Minutes to the first task.
- Integrations
- 60+ providers including Anthropic, OpenAI Codex, Copilot, Cursor, Bedrock, Vertex, Ollama and coding plans; MCP; ACP for Zed; RPC and a Node SDK; reads rules from .claude, .cursor, .codex, .cline and Copilot config.
- Data and privacy
- Code and prompts go to the model provider you choose. Sessions stay on disk. A random install ID is sent to some providers and gateways; automatic tool-issue reports need consent; OpenTelemetry export only when configured. /collab relays live sessions with frames sealed client-side.
- Cost at scale
- No paid edition. Cost is model usage or the subscriptions you log in with; round-robin keys and per-role models (a cheap smol model for subagents) help keep spend down.
- Limits
- Yolo approval and no sandbox by default; repository config, hooks and MCP servers load automatically; Alpine needs libstdc++ and libgcc for the musl binary.
What’s good
- LSP renames, debugger control and AST edits built in
- Hash-anchored edits that reject stale patches
- Worktree-isolated subagents with typed results
- 60+ providers, including subscription and coding-plan logins
What’s not
- Default approval mode runs everything without asking
- No sandbox; project hooks and extensions load on open
- Very fast release cadence and a large issue backlog
- Benchmark figures come from the author
Best for
Experienced developers who want a feature-complete terminal agent with IDE-grade code intelligence and will configure its approval settings themselves.
omp (oh-my-pi) alternatives
omp (oh-my-pi): common questions
- Is omp (oh-my-pi) worth it?
- omp is what Pi becomes when someone decides the batteries should be included, and it is among the most capable terminal agents you can run for nothing, provided you change its approval default before you trust it with a real machine. We rate it 4.2 out of 5.
- How much does omp (oh-my-pi) cost?
- Free and open source (MIT). You pay your model provider, or log in to one of the supported subscriptions and coding plans.
- What are the downsides of omp (oh-my-pi)?
- Default approval mode runs everything without asking; no sandbox; project hooks and extensions load on open; very fast release cadence and a large issue backlog; benchmark figures come from the author.
- Who is omp (oh-my-pi) best for?
- Experienced developers who want a feature-complete terminal agent with IDE-grade code intelligence and will configure its approval settings themselves.
- How much does omp (oh-my-pi) cost for a team or in production?
- No paid edition. Cost is model usage or the subscriptions you log in with; round-robin keys and per-role models (a cheap smol model for subagents) help keep spend down.
- Is omp (oh-my-pi) safe for company data?
- Code and prompts go to the model provider you choose. Sessions stay on disk. A random install ID is sent to some providers and gateways; automatic tool-issue reports need consent; OpenTelemetry export only when configured. /collab relays live sessions with frames sealed client-side.
- What are the best omp (oh-my-pi) alternatives?
- The closest alternatives we have reviewed are Pi, OpenCode and Crush.
Answers checked Nov 20, 2026.