Quick answer
Cline is the original open-source VS Code agent: simple, widely used, bring your own API key. Roo Code forked it and added modes (Code, Architect, Ask, Debug, custom) with a model per mode. Kilo Code merged ideas from both, added an orchestrator that splits tasks across sub-agents, supports JetBrains, and offers free credits through its own gateway so you can start without a key. All three are free software; you pay your model provider for tokens.
The commercial AI editors charge a subscription and hide the model bill inside it. These three do the opposite: the software is free, and every request goes straight to the provider you chose at that provider's price. That makes them cheaper for light use, more expensive for heavy agent sessions, and completely transparent either way.
Cline: the reference implementation
Cline reads and writes files, runs terminal commands with your approval, drives a browser for testing, and connects to MCP servers. It has the largest user base and the most tutorials. Its design is deliberately plain: one agent, one set of instructions, checkpoints so you can roll back. If you want to understand how these agents work, start here.
Roo Code: modes and control
Roo Code keeps everything Cline does and adds modes. Architect mode plans and is barred from editing code; Code mode implements; Ask mode answers questions without touching files; Debug mode investigates. Each mode can use a different model, so you might plan with a frontier model and implement with a cheaper one. Custom modes let teams encode their own workflows. The cost of that flexibility is configuration: Roo has many settings, and new users can spend an hour tuning before they write code.
Kilo Code: orchestration and onboarding
Kilo Code forked Roo and went in two directions. For onboarding, it provides free credits through its own gateway to hundreds of models at pass-through prices, so there is no key to paste on day one. For scale, its orchestrator mode breaks a task into subtasks and hands each to a sub-agent, which handles larger refactors than a single agent loop. It also runs in JetBrains IDEs. Being newer, it changes fast and has rougher edges.
What they cost to run
- All three: $0 for the software
- Token cost depends on the model: a session on a frontier model can cost several dollars; the same session on a budget model can cost cents
- Sub-agent orchestration in Kilo multiplies token use because each sub-agent re-reads context
- Set spending limits at your provider before running long agent tasks; none of the three will stop you from spending
Related reading
Bottom line
Install Cline if you want the simplest tool and the biggest community. Move to Roo Code when you want modes and per-task models. Try Kilo Code if you want to start without an API key or need JetBrains. They share enough DNA that switching later is painless.
