Developers can redefine their workflow with a coding agent that brings IDE precision to the terminal
You can use IDE autocompletion and debuggers directly in the terminal. References update automatically when files move, and you can attach a debugger to trace errors during execution. omp, created by Stencil Labs, is a coding agent that combines IDE integration and various execution environments with the existing Pi base.
Connecting the IDE and the agent
omp connects LSP (Language Server Protocol) to all write operations so the agent shares the information the IDE knows. When you request a file name change, re-exports, barrel files, and alias imports are updated automatically before the file moves. It also runs actual debuggers to analyze segfaults in C binaries or hangs in Go services directly with lldb, dlv, and debugpy. This allows you to inspect framework and variable states directly, moving away from debugging with print statements.
Editing strategies to boost model performance
omp uses Hashline, a content-hash-based editing method that lets the model specify anchors instead of re-entering the lines it wants to change. This prevents whitespace errors and string-not-found loops, reducing output tokens by up to 61% for the Grok 4 Fast model on the same task. It also continuously adjusts prompts for each model, raising the edit success rate for Grok Code Fast 1 from 6.7% to 68.3%. Gemini 3 Flash shows performance 5%p higher than the str_replace method, while MiniMax sees a 2.1x increase in pass rate.
Native architecture that handles everything locally
Unlike other agents that run rg, grep, find, and bash as external processes, omp links ripgrep, glob, and find inside the process. A brush bash fork and 58 command-line utilities are compiled into the builtins crate, executing without fork/exec overhead. This structure, consisting of approximately 80,000 lines of Rust core, runs as the same binary on macOS, Linux, and Windows without needing a WSL bridge. Six crates, including pi-shell, pi-natives, and pi-walker, work in parallel from the libuv pool to minimize latency in search and shell execution.
Session features for managing collaboration and memory
The /collab command puts a live session on a relay and generates a link and QR code, allowing team members to join from another terminal or browser. You can share a read-only link to observe only, or share read-write permissions to operate the same agent together. The agent remembers facts about the codebase across sessions using retain, learn, and recall commands, configured at the project scope so learning is maintained only for that repository. It also provides a monitoring feature where you designate a second model as an advisor to review every turn of the main agent and insert inline annotations.
Environment and conditions to check before adoption
omp runs on macOS, Linux, and Windows and requires bun 1.4.4 or higher. Alpine Linux users must install libstdc++ and libgcc first. Installation verifies the SHA-256 checksum to confirm binary integrity, and you can pin a specific version. Pull requests are currently open to everyone temporarily, but there was previously a vouch system, so it is best to check the latest status before contributing. It supports over 60 providers and 31 built-in tools, but some tools (e.g., github, security_scan) must be enabled in settings to be used.