AI harnesses are the software wrapped around a language model that makes it able to work. OMP – short for oh-my-pi – is one such harness: an open-source coding agent by Can Bölük that, according to its own documentation, started as a fork of Pi and extends that lean core with numerous built-in features. Licensed under MIT and available for macOS, Linux and Windows, OMP connects over 60 model providers to a tool set with LSP, debugger and sub-agent support. As of 20 September 2026.
What is OMP?
According to the repository, OMP is developed by Can Bölük, a security researcher who also operates under Stencil Labs, Inc.; the project site is omp.sh. The documentation describes OMP as a fork of Pi “by Mario Zechner”, “rewritten as a coding-first surface”. OMP installs on macOS and Linux via curl -fsSL https://omp.sh/install | sh, via Homebrew (brew install can1357/tap/omp), via Bun (bun install -g @oh-my-pi/pi-coding-agent), via Nix, or on Windows via irm https://omp.sh/install.ps1 | iex. Running omp then starts the interactive terminal interface.
The source code is published under the MIT licence in the can1357/oh-my-pi repository; vendored third-party components such as the embedded bash fork remain under their own respective licences according to the documentation.
OMP as an AI harness
OMP implements the seven building blocks far more extensively than the Pi core it descends from. The agent loop routes each turn through up to nine model roles – default, smol, slow, plan, commit, vision, task, advisor and tiny – and lets the agent raise structured follow-up questions through the ask function. Over 30 built-in tools cover file access, structured editing (edit, ast_edit), the shell (bash), code intelligence (lsp, debug), web search, browser and desktop control, and coordination (task, hub); by its own account, many of these tools run as native Rust modules in-process rather than as external programs.
On first run, according to the documentation, OMP reads existing rules, skills and MCP server configurations from the directories of other common coding agents – such as .claude, .cursor, .windsurf, .gemini, .codex, .cline, .github/copilot and .vscode – without requiring a migration. For sub-agents, the task tool fans work out into isolated working copies; the Agent Hub (shortcut Alt+A) shows status, cost and transcript for every sub-agent. A memory system with the retain, recall, reflect and memory_edit functions maintains project-scoped knowledge across sessions.
Models and model access
The documentation lists over 60 providers for OMP, grouped into frontier APIs, coding-plan subscriptions and self-hosted endpoints – relevant among others for Claude, GPT and Grok. Nine roles route different work to different models; the /model command switches the active model mid-session.
According to the documentation, custom OpenAI-compatible providers are defined in the ~/.omp/agent/models.yml file with baseUrl, api: openai-completions and apiKey; the default role can then be set in ~/.omp/agent/config.yml under modelRoles.default. This means OMP can be operated over the AI Gateway by registering the gateway as such a custom provider in that file.
Enterprise use: data protection and governance
According to the documentation, OMP runs inside the local terminal or editor and connects directly to the chosen model provider or a self-hosted endpoint; no additional cloud layer from the project itself is described. Path-bound entries (path:) let teams set allowed or blocked models per repository without touching the global configuration. The omp-stats package gives a local dashboard for a team’s own AI usage instead of vendor telemetry. Over the Agent Client Protocol in an editor such as Zed, destructive tools pause for a permission request; standalone in the terminal, no dedicated permission system is documented.
The same principle applies here: source code, file contents and prompts go to whichever model provider is selected. That belongs in your organisation’s record of processing activities beforehand, regardless of which of the more than 60 providers is used.
Orchestrating OMP with herdr
The herdr documentation lists OMP as a supported agent with lifecycle hooks once the integration is installed; through this, herdr reads both state and session information. According to the documentation, herdr can resume an OMP session via the omp --resume command after a server restart.
Where it fits: developer harness and enterprise harness
Like Pi, from which it descends, OMP is a harness for developers working in a terminal or editor, not for business users across the organisation. For that audience, CompanyGPT takes on the role of harness: tools for line-of-business systems, context from enterprise knowledge, and role-based permissions instead of a terminal and repository. The AI Gateway forms the shared cost and governance layer through which both OMP and CompanyGPT can connect to centrally managed model access. Further coding and orchestration harnesses are listed in our harness overview. We help you connect harnesses like OMP cleanly to your permission model, your model routing and your governance requirements.