Quick Verdict
OMP is the current canonical name for the open-source terminal coding agent previously described as Oh My Pi. This page evaluates v17.0.7 as of 2026-07-21. It evolved from Pi but now has a clear coding-first identity: hash-anchored edits, LSP and DAP integration, persistent Python and JavaScript execution, subagents, browser control, memory, review workflows, and ACP/RPC interfaces alongside ordinary file and shell tools.
The client is free under MIT, but operation is neither cost-free nor safe by default. Its 40+ providers include separately billed APIs, coding subscriptions, gateways, and local compatible servers. Its 32-tool surface can modify files, run commands, inspect GitHub, use SSH, and control a browser. OMP is compelling for terminal users who understand those boundaries; it is a poor choice for handing production credentials and permanent shell approval to an unreviewed model session.
Best For
OMP fits developers comfortable with Git, shells, environment variables, dependency review, and inspecting every diff. It is particularly useful for large repositories, cross-language renames, native debugging, parallel investigation, and teams that need model-provider choice on macOS, Linux, or native Windows. Cursor is easier for IDE-first editing, Claude Code offers a more integrated official terminal product, and OpenCode is another open multi-model agent worth testing.
Key Features
- Hashline edits: content-hash anchors reject stale locations instead of silently applying a patch to changed text.
- LSP and DAP: 14 LSP operations support diagnostics, references, and rename; 28 DAP operations drive debuggers such as LLDB, Delve, and debugpy.
- 32 built-in tools: files, regex and AST search, shell, persistent Python/JavaScript, SSH, subagents, browser, GitHub, image, speech, and memory; several external or sensitive tools are setting-gated.
- 40+ providers: direct APIs, coding plans, gateways, and OpenAI-compatible local servers, with role routing, fallback chains, path-scoped model restrictions, and credential rotation.
- Four entry points: interactive TUI,
omp -p, Node SDK, and RPC/ACP for terminal, automation, or compatible editors. - Review and isolation controls: prioritized review verdicts, isolated subagent worktrees, proposed AST edits, and permission prompts reduce risk but do not replace OS isolation or human review.
Use Cases
- Investigate a call chain with read, search, and LSP before applying a hashline edit and running tests.
- Drive a DAP session for a C, Go, or Python failure and inspect threads, frames, and variables.
- Fan independent research out to isolated subagents and require schema-validated results before changing code.
- Route planning and inexpensive worker tasks to different models with explicit budget and fallback limits.
- Expose OMP through ACP while retaining permission prompts for file writes and shell execution.
- Run review agents over a branch, then confirm every P0-P3 finding against the actual diff and test suite.
Pricing
| Item | Cost | Boundary |
|---|---|---|
| OMP v17.0.7 | $0 | MIT-licensed local client and source code |
| Model providers | Provider-specific | API tokens, coding subscriptions, or local compute are paid by the operator |
| Search, image, and speech services | Configuration-specific | Enabling external tools may add third-party cost and data transfer |
MIT covers OMP code, not model terms, generated-code licensing, third-party subscriptions, or an enterprise SLA. Production users should pin a version, inventory dependencies, and qualify upgrades instead of continuously executing an unpinned installer.
Pros
- Open MIT code with inspectable agent, tool, provider-routing, and extension layers.
- Hashline, LSP, AST, and DAP bring code intelligence and safer edit targeting into the terminal.
- Broad provider choice, including role-based model selection and local compatible endpoints.
- TUI, one-shot, SDK, RPC, and ACP surfaces support interactive and automated workflows.
- Worktree isolation, preview-before-apply flows, and setting-gated tools provide useful control points.
Cons
- Thirty-two tools create a large privilege surface; shell, SSH, browser, and GitHub actions need least privilege.
- Imported rules, skills, MCP servers, and extensions can inject instructions or execute code; automatic discovery does not establish trust.
- Many providers increase secret-management, cost forecasting, data-policy, and behavioral-consistency work.
- Fast releases may alter configuration and tools, so production automation should pin
v17.0.7or test every upgrade. - An open-source license supplies no SLA, response deadline, or warranty for generated changes.
Alternatives
| Tool | Better fit | Trade-off versus OMP |
|---|---|---|
| Claude Code | Anthropic’s official terminal workflow | Mature integration; less provider freedom |
| Cursor | Full AI IDE users | More visual editing; not an open terminal-first harness |
| OpenCode | Open, multi-model terminal users | Similar category with different tools, protocols, and configuration |
| GitHub Copilot | GitHub and IDE governance | Mature enterprise ecosystem; less harness-level openness |
| Goose | Another open agent architecture | Different extension model; compare edit and approval behavior on a real repository |
FAQ
Are OMP and Oh My Pi separate products?
They should not be separate directory entries. OMP is the current canonical name and command; the old oh-my-pi page is archived as a draft.
Is OMP v17.0.7 free?
The source and client are free under MIT. Cloud models, search, image, and speech services may charge separately, while local models consume operator hardware.
Which model providers does it support?
The project documents more than 40 direct APIs, coding plans, gateways, and compatible local endpoints including Ollama, LM Studio, llama.cpp, vLLM, and LiteLLM. Availability can change, so use the documentation shipped with the pinned version.
Do hash-anchored edits make changes safe?
No. They reduce misplaced edits by rejecting stale anchors. They cannot detect business-logic mistakes, malicious dependencies, unsafe commands, or excessive access.
What should be checked before installing a skill or MCP server?
Review provenance, pinned version, install scripts, dependencies, readable directories, network destinations, environment access, and write actions. Test without production credentials and disable unrelated tools.
How should a team reduce tool risk?
Use low-privilege accounts and short-lived keys, constrain the working directory, retain write and shell confirmation, isolate browser profiles, disable SSH when unnecessary, and require CI plus human review.
Bottom Line
OMP v17.0.7 is valuable because it combines editing, LSP, debugging, execution, provider routing, and subagents in an open coding harness. Those capabilities also make it substantially more dangerous than chat. Pin the version, begin in a credential-free test repository with only essential read, search, and edit tools, map provider cost and data flow, and unlock shell, browser, GitHub, or third-party extensions only after review.