Kilo Code: an open-source AI coding agent that lives in your editor

Kilo Code is an open-source AI coding extension for VS Code, the JetBrains family of IDEs and the command line. Instead of asking you to paste code into a chat window, it works inside your project: it reads the relevant files, proposes a plan, shows the exact diff it wants to apply, and runs commands only after you approve them. The vendor sums up the positioning as "one open agent for every developer workflow".

That difference matters more than raw code generation. The hard part of AI-assisted development is rarely producing a plausible function, it is making sure the change lands in the right file, respects the surrounding conventions, and stays reviewable. Kilo Code is built around that loop, and it hands model choice and key management back to the user through support for bring-your-own-key setups and a public repository.

Main capabilities

Code generation and editing. Describe the task in natural language and the agent writes new code or makes targeted edits to existing files. Changes arrive as diffs you can accept or reject one by one rather than an all-or-nothing rewrite.

Switchable modes. Code mode implements, Architect mode focuses on design and technical documentation, and Debug mode is aimed at tracing and fixing defects. Switching modes is effectively swapping the set of constraints the model works under, which keeps exploration and implementation from blurring together.

Codebase questions. Ask where a function is called, why a configuration exists, or how a module is wired up. Once the project is indexed, answers come back with file locations so you can verify them directly.

Task automation. The agent can read and write files and run terminal commands, taking over repetitive chores. Destructive actions default to manual approval, and auto-approval can be scoped by rule if you want more autonomy.

Code review. It analyses pull requests and suggests improvements, which is useful as a machine pass before a human reviewer opens the diff.

MCP integration. Through the Model Context Protocol you can expose internal tools, APIs or scripts to the agent, so company-specific operations become callable from the same panel.

Model choice. The vendor states that 500+ models are available, and you can move between fast and reasoning-heavy models per task. You can also supply your own API key.

How to use it

1. Install the extension. Open the Extensions panel in VS Code (Ctrl+Shift+X or Cmd+Shift+X), search for "Kilo Code" and install it. It is also available from Open VSX, as a manual .vsix install, and in the Cursor extension marketplace.
2. Sign in. Click the Kilo Code icon in the sidebar and complete registration or login.
3. Pick a model. Choose a model in settings, or point the extension at your own key. Starting with a smaller model is a cheap way to test the workflow.
4. Describe the task. Write the goal and the constraints, for example "add argument validation to this Python script and cover it with tests".
5. Review the proposal. The agent lists the files it wants to read, the edits it wants to make and the commands it wants to run. Approve or reject each step.
6. Iterate. Feed back on the result until the change matches your intent. Working on a separate branch keeps rollback trivial.

Where it fits

• Prototyping. Turn an idea into a runnable skeleton so you can validate the structure before polishing details.
• Maintaining unfamiliar code. Map call relationships and fill in comments or documentation when you inherit a repository.
• Pre-review screening. Run the agent over a pull request to catch obvious omissions before a colleague spends time on it.
• Learning. Ask the agent to explain an execution path and check its answer against the file locations it cites, rather than copying a solution.
• Enforcing conventions. Put style and directory rules into the agent's instructions so generated code looks like the rest of the codebase.

Limitations and things to check

• Costs scale with usage. Published plans are pay-as-you-go for individuals, USD 15 per user per month for teams, and custom pricing for enterprise. Pay-as-you-go means the bill tracks model calls, and long-context sessions add up quickly, so set a budget or limit first.
• Output still needs human review. Generated code can contain logic errors, deprecated APIs or security flaws. Anything touching authentication, money or data deletion should be read line by line before merging.
• Vendor claims are vendor claims. The 500+ model count, autocomplete and parallel agents come from official pages; what you can actually use depends on your plan, account and region.
• Data flow needs assessment. With hosted models, code snippets leave your machine. Teams working on private code should prefer bring-your-own-key, a private gateway, or the enterprise tier's audit and control features, and clear it with compliance first.
• Gaps in the source material. The collected material provided no latency or concurrency benchmarks and no independent third-party evaluation, so performance should be measured on your own repository.

FAQ

Is Kilo Code free? The software is open source, with the repository at github.com/Kilo-Org/kilocode, so you can inspect the code and the licence terms. Model usage is usually paid: the vendor lists pay-as-you-go for individuals, USD 15 per user per month for teams, and custom enterprise pricing. Check the official site for current figures.

Does it only work in VS Code? The vendor documents VS Code, JetBrains IDEs and a command-line interface, and mentions installing from the Cursor marketplace or Open VSX. Feature parity across editors may differ, so check the release notes for your editor before relying on a specific capability.

Why use this instead of pasting code into a chat window? Context and execution. The extension reads project files itself, edits at known locations, presents changes as diffs and asks for approval before running commands. That removes the copy-paste cycle and makes every change traceable.

Can it handle our private code? Technically yes, but run a compliance review first. Realistic options are a bring-your-own-key setup pointed at an approved model provider, or the enterprise tier with a private gateway, audit logs and access controls, so code does not reach an unapproved third party.