omp ─ A Terminal-Resident Coding Agent That Handles Everything from LSP Diagnostics to Live Debugging

omp is a terminal-resident coding agent with the capabilities of a full IDE packed in. Repository exploration, file editing, command execution, language-server diagnostics, and live debugging of running processes all happen inside a single session. It launched as an extended fork of “Pi,” the minimal coding agent developed by Mario Zechner, forked in late December 2025 and augmented with LSP/DAP integration, subagent support, and access to more than 60 model providers. It comes from Stencil Labs. The project is open source (MIT license) and runs on Mac, Windows, and Linux.

Key Features

Hash-line editing: Code locations are addressed by content hash, so edits rarely fail from string mismatches, and output tokens drop as well (official benchmarks report a 61% reduction with Grok 4 Fast)
LSP / DAP integration: Language-server diagnostics and symbol renaming, plus live debugging against running processes (setting breakpoints, inspecting stacks, and so on), are available to the agent as standard tools
Subagents / Agent Hub: Split a task across parallel workers in isolated worktrees and have typed results flow straight back to the parent session. A dedicated view lets you monitor each worker’s progress and step in when needed
60+ model providers: Automatic routing assigns models per role — standard, fast, advanced reasoning, planning — and fallback chains are configurable
Native Rust core: ripgrep, the shell (brush), parsers, and more run in-process across roughly 80,000 lines of Rust, avoiding fork/exec for extra speed. Windows runs natively, no WSL required
Session trees and live collaboration: Sessions can be branched, resumed, and exported, and /collab shares a link so other developers can join or observe the same session
Import your existing config: omp reads eight config formats, including Claude Code, Cursor, and Codex, so switching from another tool doesn’t mean rewriting your setup

Pricing

omp itself is free, open source software (MIT license), with no usage fee. There is no pricing page on the official site, because omp follows a BYOK (Bring Your Own Key) model: you supply and connect your own API keys or subscriptions from any of 60+ model providers such as Anthropic, OpenAI, and Google. Your actual cost depends entirely on the pricing of the model provider you choose.

Information as of September 2026. Check the official site for the latest pricing structure and supported providers.

Pros and Cons

✅ Pros

You can bring the API keys and coding plans you already pay for, so trying it out requires no new subscription
More than 31 built-in tools — LSP diagnostics, a DAP debugger, browser and desktop control, subagents — let you go from investigation to fix to verification without leaving the terminal
Free under the MIT license, and it reads other tools’ configs as-is, keeping switching costs low
Runs natively on Windows without WSL in between, making it easy to hand out across a team regardless of environment

⚠️ Cons

A third-party benchmark (Composio, published September 2026) reports that omp trails its upstream, Pi, on task success rate, speed, and token consumption alike (a single-benchmark result, so treat it as needing further verification)
It’s a large monorepo, feature-heavy and including roughly 80,000 lines of Rust core, which can be overkill for anyone who wants Pi’s minimal footprint
The AI models are not bundled, so provider API keys or subscription costs are a separate expense
A very high release cadence since its late-2025 debut brings a risk of breaking changes and instability

Comparison with Similar Services

Pi (pi-mono): omp’s upstream. A minimal setup with only four tools; Composio’s benchmark reports it beat omp on success rate, speed, and cost alike
Claude Code: Anthropic’s first-party CLI agent. omp differs in supporting far more model providers and not tying you to a single vendor
OpenAI Codex CLI: OpenAI’s first-party terminal agent. omp adds LSP diagnostics, a DAP debugger, and multi-provider routing on top
OpenCode: Another open source, multi-provider terminal agent. A close competitor in its lean toward packing in features

Who It’s For

Anyone who wants bug investigation, fixes, and test runs in an existing repository to happen in one terminal session
Anyone who wants to split audits or refactors across multiple packages among subagents and watch progress in parallel
Anyone who wants to attach a DAP debugger to a running process and live-debug alongside the agent
Anyone who wants to turn a session into a shareable link with /collab for pair programming or review with other developers

Conclusion

The deciding question comes down to how much you want to do from the terminal. If you only need the agent to investigate and edit, lighter options exist elsewhere; but if you want diagnostics, debugging, and parallel work all riding in the same session, omp covers a lot of ground. On the other hand, the benchmark results against upstream Pi and the frequent releases in a short window both warrant caution, so it’s worth checking the official documentation for the current specifications before you adopt it.