Kilo Code

Filed by the Vibe Code Weekly desk · Updated Nov 12, 2026

The story

Kilo Code bets on a single open agent that follows a developer from editor to shell to browser, with model access priced at cost. Its MIT license and its stance on model choice put it at the open, provider-neutral end of the AI coding market.

Kilo Code is built so developers are not locked into one interface or one model provider. The same agent runs as a VS Code extension, which also works in Cursor, Windsurf and other Open VSX editors, and as a plugin for JetBrains IDEs. It also runs as a command-line tool that fits into scripts and CI jobs. Agents can also be launched from a web browser, called up by mentioning @Kilo in Slack, and monitored from a phone. A desktop app for macOS, Windows and Linux, labeled beta, puts several agents next to each other in one window.

Work is split across specialized roles. A Plan agent drafts structured guidance for a complex feature before anything is written. The Code agent handles writing and refactoring with awareness of the whole codebase. A Debug agent reads error output, traces the cause and proposes a fix. Ask and Custom modes round out the set, and a code reviewer covers review workflows. Agents started locally or in the cloud can run at the same time in their own worktrees, and Kilo presents them all through one control portal.

On models, Kilo offers more than 500 options, from frontier and open-weight models to your own API keys or locally run models, and charges no markup on inference. Its Auto Model feature picks for you based on a chosen strategy: Efficient classifies the live session to control spend, Frontier sends work to the newest capable paid models, and Free sticks to free and experimental models. For teams wary of opaque tools, the code is open, so prompts, context windows and agent decisions can be inspected, and the project can be forked or self-hosted.

Kilo Code runs one open-source agent across VS Code-compatible editors, JetBrains IDEs, the shell and CI, the browser, and Slack. Its code is MIT-licensed, and models are offered at cost with zero AI inference markup, so developers can inspect the prompt and context and bring their own keys or local models.

The good

Open MIT-licensed code with visible prompts and context, and no silent model switching. Model inference at cost, with support for your own keys and local models. Distinct Plan, Code and Debug agents plus custom modes. Runs in VS Code-compatible editors, JetBrains IDEs, the shell, the browser and Slack. Auto Model routing with Efficient, Frontier and Free strategies.

The catch

Model usage still costs money unless you stick to free or local models. The desktop app carries a beta label. Getting value requires comfort with an IDE or command line.

Who should read on

Developers who switch between an IDE, the terminal and CI and want one agent across all of them. Teams that want to avoid depending on a single model provider. Engineers who want to run several coding tasks at once in separate worktrees. Organizations that need to audit or self-host their coding agent.

Who can skip it

People who want a no-code builder rather than an agent inside developer tools. Users who prefer a polished, stable desktop app today, since Kilo Desktop is in beta.

The price

Kilo Auto can be started free with no credit card required. The agent is open source under the MIT license, and models are offered at cost with zero AI inference markup; you can also bring your own keys or run local models.

Reader mail