oh-my-pi (binary name: omp) is an MIT-licensed terminal AI coding agent that has crossed 9,500 GitHub stars and 750+ forks by treating the developer's terminal as a first-class execution environment for LLM-driven coding. Built as a fork of Mario Zechner's Pi project, omp distinguishes itself with hash-anchored edits, deep LSP integration, persistent Python and JavaScript kernels, and subagent spawning, all wired into a TypeScript CLI and a Rust core. The latest release, v15.7.6, shipped on June 1, 2026. ## What omp Is omp is not a wrapper around an LLM API and a file-write tool. It is an interactive agent loop that runs in the terminal, holds long-lived runtime kernels, talks to language servers, drives a real browser, and coordinates subagents in parallel. The agent's tool surface, 32 built-in tools, is designed for engineering tasks that span multiple files, multiple processes, and multiple model providers at once. ## Hash-Anchored Edits The headline technical idea is hash-anchored edits. Instead of having the model rewrite entire files or full line ranges, omp identifies regions by content hash and applies edits at those anchors. The practical effect is dramatically reduced token usage on edit operations, since the model emits a short anchor plus a diff rather than full source. It also makes edits more robust to background changes in the file, because the anchor is content-derived rather than line-number-derived. ## LSP and AST Awareness omp integrates with Language Server Protocol implementations across major languages, which gives it real refactor operations like renames that span an entire codebase rather than per-file find-and-replace. Structural AST editing with preview-before-apply means the agent can propose changes the user inspects before they hit disk. This is closer to how a senior engineer works than the typical patch-and-pray flow of pure text-diff agents. ## Persistent Python and JavaScript Kernels omp keeps Python and JavaScript runtime kernels alive across tool calls. Variables, imported modules, and open file handles persist between steps, so the agent can experiment, inspect intermediate values, and iterate without restarting the world each time. DAP debugger support extends this to stepping through native, Go, and Python processes, which makes omp viable for debugging tasks that pure text-LLM agents cannot reach. ## Subagents and Hindsight Memory For tasks that decompose well, omp can spawn subagents in parallel with schema-validated result contracts. A hindsight memory system retains project-scoped facts between sessions so the agent does not relearn the codebase every time it is invoked. Time-traveling stream rules let the user inject constraints mid-generation without paying the context-window tax of re-prompting. ## 40+ Providers, Role-Based Routing omp supports more than 40 model providers and routes calls by role: default, smol (cheap and fast), slow (frontier reasoning), and plan (planning passes). This lets users wire a cheap model into routine tool calls while keeping a frontier model for hard reasoning steps, which is a meaningful cost lever for heavy agent use. Fourteen web search backends plus structured content extraction for code hosts and research databases round out the external-data story. ## TypeScript CLI, Rust Core The project is roughly 84% TypeScript (the CLI, SDK, and TUI) with a Rust core that runs to about 27,000 lines and contains embedded bash, regex search, and AST operations. Installation supports macOS, Linux, Windows, and Bun. The Rust core matters because it keeps the agent fast and the binary small enough to run on developer machines without becoming a heavyweight background process. ## Limitations The surface area is large. Thirty-two tools, forty-plus providers, multiple runtime kernels, browser automation, and subagent orchestration mean omp has more knobs than newcomers are likely to need, and the configuration story takes time to learn. Real browser automation via Puppeteer (with stealth mode) is powerful but also carries security and reliability implications that the user has to think about. As a fork it depends on continued alignment with the upstream Pi project for some shared mechanics, and the Rust plus TypeScript mix raises the bar for casual contributors compared to pure-Python alternatives. For developers who want a serious local agent rather than a cloud chat experience, however, omp is one of the most feature-dense options shipping today.