Kilo Code Review (2026): Pricing, Models and the Catch
An honest look at Kilo Code, the open source coding agent that sells model access at cost across VS Code, JetBrains, the CLI and the cloud.
Generate a site → Official site
The pitch, in the vendor's own words
Kilo describes itself as one open agent for every developer workflow: five hundred or more models with zero markup, frontier, open weight or your own keys, one agent everywhere you work, fully open source. That last clause is doing real work. The site says the source is MIT licensed, that you can fork, modify and self-host it, and that you can audit the prompt, the context window and the decisions the agent makes, with no silent model switching. Anyone who has watched a closed tool quietly swap models underneath them will understand why that sentence is on the homepage. The company also announces at the top of the page that Kilo has been acquired by Anaconda, which is the single biggest thing to factor into a long-term bet either way.
Zero markup, and what it really means
The pricing angle is that inference is sold at cost. Kilo states zero AI inference markup, support for bringing your own keys, and support for local models, framed as models at cost with no markup and no lock-in. Read that as a structural choice rather than a promotion: instead of bundling model spend into a monthly seat price, the tool passes provider costs through and makes its money elsewhere. For heavy users that is usually cheaper than a flat subscription, and it makes your bill track your actual usage rather than a vendor's average. It also means budgeting is on you, so put a spend alert on whichever provider you point it at before the first long session, because nothing else in the setup will stop you.
Auto Model and the Kilo Code modes
Two features shape day-to-day use. Auto Model routes requests for you across three published tiers: Efficient, described as session-aware routing for smarter spend powered by live session classification; Frontier, which goes to the latest and most capable paid models when the work needs it; and Free, which uses the best free and experimental models available. Separately, the agent ships specialised modes so you switch context without switching tools, listed as Ask, Architect, Code, Debug and Custom. Architect is for planning a feature before any code is written, Debug for tracing errors and suggesting fixes. In practice the modes matter more than the routing, because they change how you brief the agent rather than what pays for the call.
Everywhere, and whether that helps you
Kilo lists VS Code, JetBrains editors including IntelliJ IDEA, PyCharm and WebStorm, a CLI, cloud agents, Slack, a code reviewer and a gateway. Parallel isolated worktrees and an agent command center for controlling every agent from one place also appear on the feature list. Breadth like that is genuinely useful for a team spread across editors, and genuinely irrelevant if everyone on your team lives in one. Judge the surface you will actually use. The homepage also cites five million Kilo coders and more than ten trillion tokens processed monthly, which are the vendor's own numbers and worth reading as adoption signals rather than quality measurements. Adoption tells you a tool is maintained, not that it suits your work.
The claims worth checking
Models at cost
Five hundred or more models with zero inference markup, the option to bring your own keys, and local model support. Your bill then follows real usage instead of a flat seat price.
MIT licensed source
The site says you can fork, modify and self-host the agent, and inspect the prompt and the context window yourself, with no silent model switching happening behind your back.
Modes, not one prompt box
Ask, Architect, Code, Debug and Custom modes let you brief the agent differently for planning, writing and bug hunting, without leaving the editor or switching to another tool.
Now part of Anaconda
The banner at the top of the site announces the acquisition and links the full post. Factor it into any multi-year decision, in either direction, alongside the open licence.
Kilo Code next to Begin.sh
| Feature | Kilo Code | Begin.sh |
|---|---|---|
| What it is | An open source coding agent across editors, CLI and cloud | A prompt-to-project builder in the browser |
| Model choice | Five hundred plus models, your own keys, local models | Not your decision, you describe the outcome |
| Licence | MIT, per the vendor site | Hosted service |
| Output | Changes in your repository | A zip with a static site or Expo app |
| Cost shape | Inference at cost with no markup | Free to start |
| Backend and auth | Whatever you build | Not included |
| Best reader | A developer with model preferences | Someone who wants the files today |
A fair trial in four moves
- Install where you already work
Pick the one surface you use daily, VS Code, a JetBrains IDE or the CLI, and ignore the rest. Breadth is a team feature, not a first-session feature. - Start on the free routing tier
The site offers a free entry point with no credit card. Use it to judge the agent loop before you connect a paid key to anything. - Try Architect before Code
Have it plan a feature you already understand and compare the plan with your own. That is a faster quality signal than watching it write files. - Connect a key with an alert
When you move to paid models, set a spend alert at your provider first. At-cost pricing means nothing caps your usage except you.
FAQ
Is Kilo Code actually open source?
The vendor states the agent is MIT licensed and that you can fork, modify and self-host it, inspect the prompt and context window, and rely on no silent model switching. Those are strong claims and easy to verify yourself in the public repository, which is exactly what I would do before betting a team workflow on them.
What does zero markup mean for my bill?
It means inference is passed through at provider cost rather than bundled into a subscription, with the option to bring your own keys or run local models. Heavy users generally come out ahead; light users may find a flat-rate competitor simpler. Either way, set a spend limit at the provider before your first long session.
Where can I use it?
The site lists VS Code, JetBrains editors such as IntelliJ IDEA, PyCharm and WebStorm, a CLI, cloud agents, Slack, a code reviewer and a gateway, with parallel isolated worktrees for running work side by side. Most people should start with one surface rather than trying to adopt all of them at once.
What is Auto Model?
It is Kilo's routing layer, published as three tiers: Efficient for session-aware routing aimed at smarter spend, Frontier for the most capable paid models when the work demands them, and Free for the best free and experimental options. It removes the constant decision about which model to send a request to.
Does the Anaconda acquisition change anything?
The homepage announces it and links a full announcement, which is the right place to read the terms rather than a review. Practically, an MIT licence limits how much any owner can take away from existing users, but roadmap and hosted pricing are always the acquirer's to set.
What if I do not have a codebase yet?
Then a coding agent is the wrong starting point. Begin.sh takes a prompt, or a URL you want cloned, and returns a working static site or Expo app as a zip. It does not include hosting, a backend or auth, which is fine for a landing page or a prototype and a real limit for anything else.
No repo yet, just an idea
Begin.sh turns a description, or a site you want cloned, into a working static project you download and own. No editor plugin, no model keys, free to start.
Generate a site →