Bullet Review (2026): Pricing, Features & Honest Verdict
Bullet is a CLI coding agent built for speed that runs on your existing Claude Code or Codex subscription and claims to finish tasks 30 to 60 percent faster. Best for: developers who already pay for a coding agent and are losing time to latency. Price: free, no subscription, you bring your own model access. Rating: 7/10.
What is Bullet?
Bullet is a command-line coding agent that attacks a problem most agent vendors ignore: waiting. The pitch on the site is “a really fast coding agent” that focuses on “keeping up with you”. It installs as an npm global package, it runs on macOS and Linux, and it does not sell you inference. You connect your existing Claude Code or Codex subscription over OAuth, or drop in API keys, or point it at a model running on your own machine. Bullet is the harness, not the model.
We found it on Product Hunt, where it launched at number five with 190 upvotes and the blunt tagline “30-60% faster than Claude Code and Codex”. It comes out of Y Combinator’s P26 batch, built by Adi Kulkarni as CEO and Alex Sima as CTO, with a team of Yale computer science graduates and former AppLovin and Citadel engineers. The headline claim is 95.8 percent on SWE-bench Verified at 119 seconds per task, which would put it in the top three on that benchmark. The current release is 1.3.26 and access is still gated as a private beta.
What are Bullet’s key features?
Smart model routing per prompt
Bullet picks the model and reasoning level for each individual prompt rather than running everything through one setting. The framing on the site is “route straightforward work to fast models, escalate only when the mission demands it”. This is the single biggest lever on perceived agent speed and almost nobody exposes it well. Most agents run maximum reasoning on a request to rename a variable, which costs you 20 seconds and a pile of tokens for no benefit. Routing that to a fast model is free latency. It is also the feature most likely to occasionally pick wrong, and there is no published detail on how you override the router when it does.
Targeted code search instead of repo embedding
Bullet does not embed your whole repository. It searches for the code that matters for the current task and loads only that. The demo on the site shows a task completing with 8 percent repository context. Skipping the embedding step removes the indexing wait when you open a new project and removes a stale-index failure mode entirely, and a smaller context window is both cheaper and, in practice, more accurate. The trade is real though: targeted search is only as good as its retrieval, and on a task that genuinely needs cross-cutting knowledge of a large codebase, a well-built index will beat it.
Parallel tool execution
Independent tool calls run at the same time instead of one after another. If the agent needs to read four files and run two commands and none of them depend on each other, they all go at once. The site shows three operations completing in 2.31 seconds. This is unglamorous engineering and it is where a lot of the claimed 30 to 60 percent almost certainly comes from, because serial tool calls are how most agents burn wall-clock time doing nothing.
Bring your own everything
OAuth into a Claude Code or Codex subscription you already pay for, use raw API keys, or run an on-device model. Bullet takes no cut and adds no markup. For anyone who has already committed to a $20 or $100 a month agent plan, this means trying Bullet costs nothing but the install, and it means your existing rate limits and quotas carry over unchanged.
A one-command install
Setup is npm install -g @trybullet/cli and you need Node 18 or newer. That is the whole thing. There is no daemon, no indexing pass to sit through before you can ask the first question, and no project configuration step, which is a direct consequence of the decision not to embed the repository. Compare that to agents that want ten minutes to vectorise a large monorepo before you learn whether you like them. Access is still gated behind a private beta signup, so the install command alone will not get you running today, and that is the main friction between reading about Bullet and testing the claims yourself.
How much does Bullet cost?
Bullet is free. The site is explicit: “free to use, no subscriptions”. There is no paid tier, no seat pricing and no usage metering. Your only cost is the model access you already have, whether that is a Claude or Codex subscription, pay-as-you-go API keys, or a local model that costs you nothing but electricity.
Against the field that is unbeatable on price and slightly worrying on sustainability. Cursor runs $20 a month, Claude Code is bundled into Anthropic plans from $20 to $100 a month, and Codex sits inside ChatGPT subscriptions on similar terms. Bullet is not competing with those on price because it sits on top of them. What you should read into free is that this is a YC company in private beta buying adoption, and a pricing page will appear at some point. That is not a criticism of the product, but do not build a team workflow on the assumption that free is permanent. The fact that it uses your own credentials at least means a future price change leaves you with your model access intact.
| Plan | Price | Plan Features | Best For |
|---|---|---|---|
| Free | $0 | Full CLI, macOS and Linux, Node 18+. Bring your own Claude Code or Codex subscription via OAuth, raw API keys, or an on-device model. No markup on inference. Private beta access required. | Developers who already pay for a coding agent subscription |
Who is Bullet best for?
Use Bullet if you already pay for Claude Code or Codex, work on a large repository, and find yourself waiting on the agent often enough to notice. The install is one npm command and the switching cost is close to zero because it uses the subscription you already have. Developers who run agents in a tight loop, dozens of short tasks an hour rather than one long autonomous run, will feel the parallel execution and the routing most clearly. It also suits anyone who dislikes repository embedding on principle, whether for indexing time or for not wanting the whole codebase vectorised.
Skip Bullet if you are on Windows, because it ships for macOS and Linux only. Skip it if you need a GUI, since this is a terminal tool with no editor integration story published yet. And skip it if you are choosing your first coding agent, because Bullet is a speed layer over an agent subscription you are expected to already own, and evaluating a private-beta harness at version 1.3.26 is not where a newcomer should start. One more caution: the 95.8 percent SWE-bench Verified figure and the 30 to 60 percent speed claim are both self-reported by the vendor with no independent verification we could find. Treat them as marketing until a third party reproduces them, and judge the tool on your own repository for a week instead.
Best Bullet alternatives
Claude Code is the thing Bullet is measured against, and it is the mature option: better documented, cross-platform, backed by Anthropic, with a large plugin and skill ecosystem. It is slower on the specific axes Bullet optimises. Bundled with Anthropic plans from $20 a month.
OpenAI Codex covers similar ground from the other major vendor, with strong cloud-delegated task support and tight ChatGPT integration. It has the same serial-tool-call latency problem Bullet is built to fix. Included in ChatGPT Plus and Pro plans, from $20 a month.
Aider is the open-source terminal alternative, model-agnostic, works with any API key, and has years of real-world use behind it. It is less automated than Bullet and does no smart routing, but nothing about it is a private beta and you can read the code. Free, you pay only for tokens.
Final verdict: is Bullet worth it?
Bullet is doing something specific and doing it in the right place. Parallel tool execution, per-prompt model routing and targeted search instead of full-repo embedding are three real engineering answers to why agents feel slow, not three features invented for a launch post. Because it runs on your existing subscription and costs nothing, the evaluation is close to free: install it, run a week of your normal work through it, and see whether the waiting goes away. That is a much better test than any benchmark number, and it costs you nothing but an npm install and a beta invite. The measurable outcome to watch for is not whether tasks feel snappier, which is easy to imagine, but whether you stop context-switching away while the agent thinks. If you keep alt-tabbing to Slack mid-task, the speed gain was not big enough to change how you work.
What holds it back from a higher score is everything around the product rather than the product itself. It is a private beta at version 1.3.26 from a company months old, the speed and benchmark claims are entirely self-reported, there is no Windows support, and free with no published business model means a pricing decision is coming. None of that makes it a bad tool. It makes it an early one. We rate it 7/10 and recommend it as a second agent worth trying today, not as the one you standardise your team on this quarter.
Bullet Pros & Cons
What We Like
- Free with no subscription, running on the Claude Code or Codex plan you already pay for
- Per-prompt model routing sends trivial work to fast models instead of burning reasoning on a rename
- Targeted code search instead of repo embedding removes the indexing wait and the stale-index failure mode
- Parallel tool execution, where most agents burn wall-clock time running independent calls serially
- One-command npm install with no daemon, no config step and no vectorisation pass
What Could Be Better
- The 95.8% SWE-bench Verified score and 30-60% speed claim are self-reported with no independent verification
- macOS and Linux only, with no Windows support
- Still gated behind a private beta at version 1.3.26, so you cannot just install and go
- Free with no published business model means a pricing decision is coming
Bullet FAQ
What is Bullet?
Bullet is a command-line AI coding agent built for speed. It routes each prompt to an appropriate model and reasoning level, runs independent tool calls in parallel, and uses targeted code search instead of embedding your whole repository. It runs on your existing Claude Code or Codex subscription, API keys, or an on-device model. It comes out of Y Combinator P26 and is at version 1.3.26.
How much does Bullet cost?
Bullet is free with no subscription, no seat pricing and no usage metering. The site states plainly: free to use, no subscriptions. Your only cost is the model access you already have, whether that is a Claude or Codex plan from $20 a month, pay-as-you-go API keys, or a local model. Bullet takes no cut and adds no markup on inference.
Is Bullet worth it?
Worth trying, yes. Because it runs on a subscription you already pay for and costs nothing, evaluating it is close to free. Parallel tool execution, per-prompt routing and targeted search are real fixes for real latency problems. We rate it 7/10, held back by a private beta, self-reported benchmarks, no Windows build and no published business model.
What are the best Bullet alternatives?
Claude Code is the mature, cross-platform, well-documented option Bullet measures itself against, bundled with Anthropic plans from $20 a month. OpenAI Codex covers similar ground with tight ChatGPT integration, included in Plus and Pro plans from $20 a month. Aider is the open-source terminal alternative, model-agnostic and free, though it does no smart routing.
Does Bullet offer a free plan?
Bullet is entirely free. There is no paid tier at all. You connect your own Claude Code or Codex subscription over OAuth, supply API keys, or point it at an on-device model, and Bullet charges nothing on top. Access is currently gated behind a private beta signup, so free does not yet mean instantly available.
Who is Bullet best for?
Developers who already pay for Claude Code or Codex, work on a large repository, and notice themselves waiting on the agent. It suits people running dozens of short tasks an hour rather than one long autonomous run. Skip it if you are on Windows, if you need editor integration, or if this would be your first coding agent.






