Kilo Code: the open source AI coding agent that shows you the meter

Picture a Friday afternoon. A coding agent has just opened a pull request in your repo, the CI badge is green, and your teammate shrugs and says, "it works." You feel good about it. Then someone in the channel asks the question that actually matters: which provider did that task run on, and what did it cost? If your agent is Kilo Code, you can answer both. That is the whole pitch in one sentence.

Kilo Code is an open source AI coding agent that lives in the tools you already use. It ships as a VS Code extension, a JetBrains plugin, and a command line agent, and the team behind it also runs cloud agents and automated code reviews from the same account. The code is MIT licensed, so you can read every prompt the agent sends and check what it decided to do. The GitHub repo sits at around 27,000 stars, and Kilo's own site says it has been handling well over 10 trillion tokens a month.

Where it actually lives

Most coding agents pick a surface and camp there. Kilo takes the same agent and puts it in VS Code, the JetBrains IDEs, a terminal, and the cloud. The VS Code extension is the front door: install it, and you get access to 500 plus models through the built in gateway, with no API key needed to start. If you live in IntelliJ, PyCharm, WebStorm, or another JetBrains tool, there is a plugin in the JetBrains marketplace. The CLI version is a fork of the open source project OpenCode, and it runs in any project directory, including CI.

That last part matters more than it sounds. When an agent can run headless in a pipeline, "AI helped me code" quietly becomes "AI codes in our build system." Kilo's autonomous mode is one command with a flag that turns off the permission prompts, and the docs are honest about the tradeoff: use it in trusted environments where nobody is sitting at the keyboard. It is a small detail, but it is the difference between a demo agent and a teammate.

You do not have to pick a lane forever, either. Kilo is pitched as one agent that moves with your work: start in the editor, hand a long running task to a cloud agent, and come back to review. The cloud side also includes automated code reviews you can attach to pull requests, which is a nice way to try an agent without giving it the wheel on your main branch.

The five agents inside

Kilo organizes work with specialized agent modes you switch between, like tabs in a browser. The default Code mode writes and refactors code from plain language. Plan mode designs the approach before anything gets written. Ask mode answers questions about your codebase without touching a single file. Debug mode chases errors and traces issues. Review mode looks over your changes for performance, security, and test gaps. You can also build custom modes for the specific jobs your team repeats.

Underneath the modes, the agent does the things you would expect from a modern coding copilot: generating code across multiple files, inline autocomplete that you accept with the tab key, self checking so it reviews and corrects its own work, running terminal commands, driving a browser, and wiring up extra tools through a built in MCP marketplace. MCP is a standard for handing an agent extra capabilities, and having a marketplace inside the tool means you are less likely to waste an evening configuring it by hand.

The pricing page reads like a challenge

Here is where Kilo stops sounding like every other AI tool. The pricing page separates three things: the platform, the model inference, and cloud compute. The platform is free for individuals, and the open source extensions work with any model. Inference is where most vendors quietly add a margin, and Kilo's gateway advertises exact provider rates with zero markup across 500 plus models. If you would rather bring your own keys from OpenAI, Anthropic, Google, AWS Bedrock, or somewhere else, go ahead. If you want to run local models through Ollama or LM Studio, the extension supports that too. There is also a Kilo Pass, a monthly credit subscription with up to 50 percent bonus credits, for people who would rather budget a flat number.

Teams pay a flat per user fee for collaboration features like shared agent modes and analytics, and enterprise adds SSO, audit logs, and a private gateway. The honest way to think about the numbers is this: the tool itself will not cost you money. What it costs is the models you run it on, and Kilo refuses to sit in the middle and take a slice.

There is also Auto Model, which routes each task to a model strategy based on the session. You pick a tier, something like a frontier option for hard problems, an efficient option for everyday work, or a free tier that routes to free models, and the routing handles the rest. For a team that has ever paid frontier model prices for a "rename this variable" task, that single setting is probably worth the install.

Why Anaconda came calling

In mid July 2026, Anaconda announced it was acquiring Kilo Code. The package data company has built its name on Python package management, and its public story frames the deal as a bet on developer control over AI. The post argues that most companies do not actually know where their AI data goes, because builders are running agents on personal accounts with their own keys, and a Kilo style setup with an auditable, self hostable option puts the data back on terms the company can understand. Kilo's own homepage now carries the acquisition notice above the fold, which tells you the deal is real and settled.

It is an interesting combination. Anaconda already knows the pain of "who is shipping what version of this package." Kilo knows the pain of "which model touched this code, and who is watching the meter." Both are basically audits. The open source license means the acquisition does not change the rules for anyone self hosting today, and that is usually the part that worries a community most.

Who should try it first

If you code in VS Code or JetBrains and want an agent you can actually audit, Kilo is one of the cleanest open source options right now. The MIT license, the readable source, and the zero markup gateway make it a reasonable default to compare against whatever agent you already run. If your team is nervous about data leaving the building, the local model and self hosting story is worth a test week. And if your pipeline is hungry for automation, the CLI in CI is the feature that keeps getting overlooked: a green build that fixed its own test failures is a strange and wonderful thing to catch a Friday afternoon.

None of this is a free lunch. The agent still writes code you should review, the model bill is still yours if you use the paid tiers, and a self hosted MIT project is only as strong as the people maintaining it. But in a market full of closed agents and opaque pricing, a tool that hands you the source, the meter, and the keys is the kind of offer that is hard to argue with. Give the free tier a go on a repo you trust, watch what it does, and let the code tell you the rest.