Open-source desktop app to run, serve and fine-tune text, image, video and audio models entirely on your own machine.
Summary
A coding agent tuned for latency: routes easy work to fast models, pulls only the code it needs, and fans out tool calls in parallel.
Screenshots
Description
Bullet is a coding agent built around a single complaint — that most agents are slow not because the model is slow, but because the harness wastes time. It attacks that with three protocols rather than a bigger context window.
Route/Escalate sends straightforward work to fast models and only escalates to a heavyweight model when the task actually demands it, so trivial edits do not pay frontier-model latency. Search/Acquire skips whole-repository embedding and retrieves only the code relevant to the task — the project's demo runs on roughly 8% of repository context. Parallel Execution issues independent tool calls simultaneously instead of queuing them, which is where much of the wall-clock saving comes from: operations that normally take minutes finish in seconds.
The speed pitch is backed by a benchmark rather than a vibe. The team reports 95.8% on SWE-bench Verified across all 500 verified instances, with a repository-level breakdown published on 30 July 2026; the methodology writeup is on their blog and worth reading before treating the number as comparable to others.
Bullet ships both as a downloadable GUI and as a CLI installed from npm, currently at version 1.3.26. It runs on macOS and Linux and needs Node 18 or later.
The project is Y Combinator-backed and currently in a free private beta — no subscription, and you do not need to bring your own API key to start. Since pricing has not been announced, treat the current cost as a beta condition rather than a permanent one.
Aimed at developers who already work with a coding agent daily and find the waiting, rather than the output quality, to be the thing that breaks their flow.
Reviews
Similar App Suggestions
Spotify's vendor-neutral workspace for running dozens of Claude Code, Codex and Gemini CLI sessions in parallel with shared context.
Run agent evals and private benchmarks in sandboxed cloud environments — compare Claude Code, Codex, Cursor and Copilot on the same real tasks.
An open-source proxy that sits between your coding agent and the model, compressing tool schemas, file reads and stale history to cut token bills.
A GPU-accelerated code editor written in Rust, with collaborative editing and coding agents built into the core.