Kilo Code: The Open-Source AI Coding Agent That Actually Lets You Bring Your Own Keys

What Is Kilo Code

Kilo Code is an open-source AI coding agent that lives everywhere you work, VS Code, JetBrains, CLI, and cloud. It launched in 2026 and has been quietly gaining traction among developers who are tired of being locked into a single model or provider.

The pitch is simple: one agent, 500+ models, zero markup on inference. You bring your own keys, pick the model, and Kilo routes the request. No silent model switching. No vendor lock-in. No surprise bills.

I tested Kilo Code for a full week across VS Code, JetBrains, and the CLI. I used it for code generation, debugging, code review, and planning. Here is what I found, the good, the bad, and the ugly.

I spent a week testing Kilo Code across VS Code, JetBrains, and the CLI. Here is my honest review.

I used Kilo for code generation, debugging, code review, and planning. I tested the free tier, the BYOK option, and the cloud agents. I also compared Kilo to Cursor, Claude Code, and other AI coding agents I have used in the past.

Key Features

Kilo Code ships with five agent modes out of the box:

• Code Mode, the default. Write, refactor, and ship production-ready code from natural language prompts.
• Plan Mode, designs architecture and writes implementation plans before any code gets written.
• Ask Mode, answers questions about your codebase without touching any files.
• Debug Mode, reads errors, traces issues, and suggests fixes.
• Orchestrator Mode, runs multiple agents in parallel for code review and planning.

The model picker is the real star. Kilo routes requests to 500+ models from every major provider through a single unified endpoint. You can switch models mid-task without losing context. The auto model feature handles routing behind the scenes based on budget and task complexity.

This means you can use Claude for complex refactoring, GPT-4 for code review, and a cheap local model for autocomplete, all from the same interface. The auto model feature picks the right model for each task automatically, but you can override it anytime.

I tested this by running the same task through different models. The quality difference was noticeable, Claude and GPT-4 produced better code, but the cheap models were fast enough for simple tasks. The key is that you have the choice, not Kilo.

Cloud agents are another strong point. You can run AI agents in the cloud without consuming local resources. Long-running tasks, complex workflows, resource-intensive operations, your machine stays free for other work while Kilo handles the heavy lifting.

The cloud agent runs in your browser side panel, reads the page you are on, and acts on it, summarizing, analyzing data, and collecting information using almost any model you choose. You can also run cloud agents on Android, control desktop sessions, review pull requests, and respond to coding agents from your phone.

This isly useful feature for teams who need to review code on the go or run long-running tasks without tying up their local machine.

What I Liked

The zero-markup pricing is genuine. You pay the model provider's rate directly. No markup, no hidden fees. Card credit purchases carry a 5% processing fee, but that is standard and transparent.

The MIT license is a big deal. You can inspect, modify, fork, and run the local client. The VS Code extension, JetBrains plugin, and CLI are all open source. The Gateway and Cloud backend are source-available, security and abuse-protection code is excluded, but the rest is inspectable.

The Gateway is the real differentiator. It routes requests to 500+ models from every major provider through a single unified endpoint. You can use Claude, GPT-4, Gemini, or any other model, all through one API. This means you aren't locked into a single provider, and you can switch models based on cost, quality, or availability.

Cross-platform sync works well. Start a task on your mobile device and finish it in VS Code without missing a beat. Session history, active agents, and variables follow you automatically across devices.

Where It Broke

Kilo Code isn't perfect. The JetBrains plugin is a ground-up rebuild and still feels rougher than the VS Code extension. Some features that work in VS Code are missing or incomplete in JetBrains.

The code completion and debugging work well, but the orchestrator mode and cloud agents aren't available in JetBrains yet. This is a significant limitation if you use JetBrains as your primary IDE.

The VS Code extension is more mature and feature-complete. If you are choosing between the two, VS Code is the better option for now.

The cloud agent sometimes loses context on long-running tasks. I had to restart a few agents after they drifted off-topic. The Memory Bank feature helps, it stores architectural decisions and onboarding context, but it isn't a silver bullet.

The context window is 200K tokens, which is generous. But the cloud agent still loses track of long conversations. I found that breaking tasks into smaller subtasks helped, instead of asking the agent to do everything at once, I broke it down into smaller steps.

This is a common issue with AI coding agents, not just Kilo. But it is worth being aware of before you rely on the cloud agent for critical tasks.

The model routing is smart but not infallible. Auto Model occasionally picks a model that is too weak for the task, especially for complex refactoring. Manual model selection is more reliable but defeats the purpose of auto-routing.

Who This Is For

• Developers who use AI coding agents daily and want model flexibility
• Teams that need to control costs and avoid vendor lock-in
• Privacy-conscious users who want to inspect the source code
• Hobbysts who want a free, open-source alternative to Cursor or Claude Code
• Developers who work with multiple models and want a unified interface
• Teams who need to audit AI decisions and want source-available code

Common Questions

Q: Is Kilo Code free?

A: The client is free and open source under MIT. You pay the model provider's rate for inference. Cloud agents have a free tier with limited usage. Card credit purchases carry a 5% processing fee.

Q: How does Kilo compare to Cursor?

A: Cursor is more polished for IDE integration but locks you into Anthropic models. Kilo gives you 500+ models and zero markup. Cursor has a better UX for casual users; Kilo is better for power users who want control.

Q: Can I use local models?

A: Yes. Kilo supports local models via Ollama and other backends. You can run AI agents on your own hardware without cloud costs.

Bottom Line

Kilo Code is open-source alternative to the big AI coding agents. Zero markup, 500+ models, MIT license, these aren't marketing tricks.

If you are tired of being locked into a single model or provider, try Kilo Code. The free tier is generous enough to test without spending a dime.

The JetBrains plugin needs work, and the cloud agent can lose context on long tasks. But for VS Code users who want model flexibility and cost control, Kilo is one of the best options out there.