Bullet
A coding agent built for speed, claiming 30-60% faster runs than Claude Code and Codex
Overview
Bullet's whole premise is that Claude Code and Codex are capable but slow, and that a leaner orchestration layer, routing easy work to cheap fast models, searching for only the relevant code instead of embedding a whole repo, and batching turns to cut round trips, can close most of the gap without a bigger model. The founders (Adi and Alex) built it after burning hours waiting on their own agent runs, and it ships as a free download for Mac and Linux plus an npm CLI, with no subscription: you bring your own OpenAI or Anthropic API key and pay the model provider directly.
It launched on Hacker News this week (121 points, 89 comments) claiming runs 30-60% faster than Claude Code and Codex. Bullet's own benchmark post also claims 95.8% on SWE-bench Verified, a number Bullet published itself, not an independently audited result, and Hacker News commenters immediately pushed back on it, noting SWE-bench Verified is a widely gamed, near-saturated benchmark and that the company excluded its own routing logic from the number for a "fairer" comparison. Commenters also flagged that its downloadable binaries are unsigned and request elevated system access without published source code, a real concern for a tool with this much access to your machine and codebase.
Key Features
Model Routing: Sends straightforward tasks to faster, cheaper models and escalates to a stronger model only when the task actually needs it.
Targeted Code Search: Searches the repo for relevant code instead of embedding the whole codebase into context.
Parallel Execution: Runs independent operations at the same time instead of queuing them one after another.
Bring-Your-Own API Key: Free to use, no Bullet subscription, connects to your own OpenAI, Anthropic or other model API key.
Desktop App Plus CLI: Available as a Mac/Linux desktop download and as an npm-installable CLI (npm install -g @trybullet/cli).
Pricing
Starting price
Free (bring your own API key). Bullet itself has no subscription; you pay your model provider, OpenAI, Anthropic, etc., directly for usage. No flat per-tier price exists to report.
Bullet is free to use with no subscription tiers. You connect your own API key from a model provider, OpenAI, Anthropic and others are supported, and pay that provider directly for usage; Bullet itself charges nothing. The team has mentioned a guest mode in development and is reportedly considering open-sourcing the project, but neither is live yet.
Disclaimer: Bullet is days old as a public launch. Its cost model (bring-your-own API key, no Bullet fee) is confirmed on its own site and its Hacker News launch thread as of 2026-09-08, but could change as the product matures. Note: our research found Bullet publicly downloadable (Mac/Linux app plus npm CLI, no waitlist) rather than in a closed private beta.
Pros
YC-Backed, Real And Downloadable: Not vaporware or a waitlist, it's a working Mac/Linux desktop app and npm CLI you can install today.
Free To Use: No subscription, you bring your own model API key and only pay the model provider for what you use.
Real Engineering Ideas Behind It: Model routing, targeted code search instead of full-repo embedding, and parallel execution are genuine, explainable approaches to cutting agent latency, not just marketing.
Founders Engaging Publicly: The team answered detailed technical questions, including routing methodology and benchmark exclusions, directly in the Hacker News comments rather than avoiding scrutiny.
Cons
Its Headline Benchmark Is Self-Reported: The 95.8% SWE-bench Verified figure is Bullet's own claim, not an independently audited result, on a benchmark widely considered near-saturated and easy to game. Treat it as a marketing number, not an established fact.
A Real Security Question, Unanswered: Hacker News commenters flagged that Bullet's downloadable binaries are unsigned and request elevated system access, with no published source code to verify what they actually do. That's a serious ask for a tool that touches your codebase and machine.
Brand New, No Track Record: Launched this week. There is no independent long-term usage data, and multiple HN commenters questioned whether the business model (free tool, bring-your-own API key) is viable long-term.
No Monetization Path For Us: No affiliate or referral program exists, so this listing carries no revenue for AIwithKay, it's coverage for its own sake.
Name Collision Risk: "Bullet" is also the name of an unrelated, older Notion-to-website builder at bullet.so. This listing is specifically about codewithbullet.com, a different product.
What Makes It Unique
Bullet doesn't claim a smarter model, it claims a leaner harness around existing models: routing, targeted search and parallel execution to cut the time an agent spends waiting, rather than trying to out-reason Claude Code or Codex on raw capability.
Kay Score
5.5
/ 10
Tool Information
Pricing
Free (bring your own API key). Bullet itself has no subscription; you pay your model provider, OpenAI, Anthropic, etc., directly for usage. No flat per-tier price exists to report.
Category
Coding & Development
Platform
Web / iOS / Android
Last Updated
