Cline, Roo Code and Kilo Code vs Claude Code

Cline and Kilo Code are open-source agents that run in your editor with whichever model you bring; Claude Code is Anthropic’s own agent for Claude models. How they differ on models, modes, approvals, checkpoints, rules files and MCP, what happened to Roo Code, and who each suits.

7 min read

Cline and Kilo Code are open-source coding agents that began as VS Code extensions and let you choose the model: your own API key for Anthropic, OpenAI, Google and dozens of other providers, a gateway, or a local model. Claude Code is Anthropic’s agent, built for Claude models and signed in with a Claude or Console account. All three read the codebase, edit files, run commands and use MCP. The practical differences are model choice, how each asks for approval, how you undo, and which rules files each reads. And one name has changed status: Roo Code, the fork between the two, was shut down in May 2026.

How the three are related

The lineage explains why they look alike. Roo Code started as a fork of Cline. Kilo Code started as a fork of Roo Code; Kilo’s own announcement (opens in a new tab) from April 2025 says it forked Roo Code, “which itself was a fork of Cline”, and set out to carry features from both. So the modes, the approval toggles and the checkpoint idea have a common ancestor.

Since then they have separated. The Roo Code docs (opens in a new tab) now open with a notice that the extension was shut down on 15 May, pointing users to Cline and to a community fork, and the Roo Code repository on GitHub is archived. If you still run Roo Code, it will get no more fixes; plan the move.

Kilo Code rebuilt its VS Code extension in April 2026 on an open-source core it shares with its CLI and cloud agents, and the Kilo Code repository (opens in a new tab) says the Kilo CLI is a fork of OpenCode. The family resemblance remains; the code bases no longer move together.

Where each runs

  • Cline: a VS Code extension, a JetBrains plugin, a CLI (npm i -g cline), a desktop app, and other editors through the Agent Client Protocol. Licensed under Apache 2.0.
  • Kilo Code: a VS Code extension, a JetBrains plugin, a CLI (npm install -g @kilocode/cli) and cloud agents. Licensed under MIT.
  • Claude Code: a terminal app, VS Code and JetBrains extensions, the Claude desktop app and the web. Distributed as a binary under Anthropic’s terms. The surfaces are compared in Claude Code desktop vs CLI vs VS Code.

Models: bring your own, or Claude

Cline lists more than 30 providers: Anthropic, OpenAI, Google Gemini, AWS Bedrock, OpenRouter, any OpenAI-compatible endpoint, local models through Ollama or LM Studio, and its own pay-as-you-go provider. You can set one model for planning and another for acting. Kilo Code offers hundreds of models through its Kilo Gateway, with your own keys for Anthropic, OpenAI, Google and others if you prefer. You pay the provider, so cost tracks the model you pick.

Claude Code runs Claude models through an Anthropic account, Amazon Bedrock, Google Cloud or Microsoft Foundry. It can be pointed at other endpoints, but Anthropic does not support that; the trade-offs are in Claude Code with Ollama. If you want to change model per task, the extensions are built for it and Claude Code is not.

There is an overlap. Cline’s Anthropic provider page (opens in a new tab) offers an API key, or a Claude Code option that runs your installed claude binary so a Pro or Max subscription is used instead of API billing, with limits on streaming, images and prompt caching. Anthropic’s legal and compliance page says subscription sign-in is designed for ordinary use of Claude Code and Anthropic’s own apps, and that developers building other products should use API keys. Read both before you rely on that path.

Modes: Plan and Act, agents, permission modes

Cline has two modes. In the words of its Plan and Act guide (opens in a new tab), Plan mode can read the codebase and discuss strategy but cannot modify files or run commands; Act mode carries the same conversation forward and does the work. /deep-planning runs a longer planning pass that asks clarifying questions first.

Kilo Code calls its modes agents now: Code, the default, Ask, which is read-only, Plan, which may write only plan files under .kilo/plans/, and Debug. Its Orchestrator is deprecated because the other agents can delegate to subagents themselves, and you can define your own agents with their own tools and permissions. Roo Code’s Code, Architect, Ask, Debug and custom modes live on in Kilo in this form.

Claude Code folds planning into its permission modes: plan explores and writes a plan without editing source, and the others decide how much needs your approval. Custom roles are subagents in .claude/agents/. How plan mode works is in Claude Code plan mode.

Approvals and checkpoints

  • Cline asks for approval on every action by default. Auto Approve switches on categories: read or edit project files, run commands the model marks safe, use the browser, use MCP. YOLO mode approves everything, and Cline’s own docs call it dangerous.
  • Cline checkpoints: its checkpoints guide says it commits your files to a shadow Git repository after each tool use, separate from your own history. You can compare, restore files, restore the conversation, or both.
  • Kilo Code sets allow, ask or deny per tool, with patterns, in each agent’s configuration, so bash can ask while uv * is allowed. Snapshots, formerly checkpoints, are Git-based and let you revert to just before any message you sent.
  • Claude Code now starts terminal and VS Code sessions in auto, where a classifier approves most actions; Manual asks before most edits and commands, and acceptEdits sits between the two. It checkpoints before every prompt, and /rewind restores files it edited, though not changes made by shell commands.

The effect is similar: approve a lot at first, loosen as you learn what the agent does, and keep an undo within reach. Cline’s shadow repository also captures files Git does not track, which is worth knowing before you restore.

Rules files

Cline’s rules documentation (opens in a new tab) reads Markdown rules from .clinerules/ or .cline/rules/ in the workspace, global rules from a Cline Rules folder in your Documents directory, and, for compatibility, .cursorrules, .windsurfrules and AGENTS.md. Each rule can be toggled on or off. Kilo Code reads AGENTS.md, rules listed under instructions in kilo.jsonc, and the older .kilocode/rules/ folder. Claude Code reads CLAUDE.md, and AGENTS.md only when no CLAUDE.md exists.

The file all three can share is AGENTS.md. Put the common rules there and keep tool-specific extras in each tool’s own file; AI context files compared has the precedence rules.

MCP

All three are MCP clients for local and remote servers. Cline keeps servers in an mcpServers JSON file, opened from its MCP Servers panel, or ~/.cline/mcp.json for the CLI; set "type": "streamableHttp" for a remote server, because leaving it out falls back to the older SSE transport. Kilo Code uses an mcp key in kilo.jsonc, with local and remote types and OAuth for remote servers. Claude Code uses claude mcp add and a project .mcp.json.

Who each suits

  • You want to choose the model per task, use a local model, or pay a provider directly: Cline or Kilo Code.
  • You want to read the agent’s code or run it where a proprietary binary is not allowed: Cline (Apache 2.0) or Kilo Code (MIT).
  • You want one vendor that makes the model and the agent, with support and terms from the same place: Claude Code.
  • You work mostly in a terminal and script your tools: Claude Code, or the Cline and Kilo CLIs.
  • You are on Roo Code: move to Cline, Kilo Code or another maintained agent; Roo Code is no longer maintained.

Keeping the work on one board

Teams that let people choose their agent end up with work scattered across chat panels. fenbs puts it back in one list: any of these agents can connect to https://fenbs.ai/api/mcp and read, create and move tasks under your role, and every change is recorded in History under the assistant’s name, on your behalf. If an agent cannot complete a browser sign-in, issue a token under Settings with a name, the scopes it needs and an optional expiry, and revoke it when you are done.

Cline MCP settings (mcpServers)
{
  "mcpServers": {
    "fenbs": {
      "type": "streamableHttp",
      "url": "https://fenbs.ai/api/mcp",
      "headers": { "Authorization": "Bearer <token from fenbs Settings>" }
    }
  }
}

Related

Connect Claude Code: the Claude Code integration. The tools every agent gets: the MCP docs. Other comparisons: Gemini CLI vs Claude Code vs Codex CLI, OpenCode vs Claude Code and Claude Code vs Cursor.

Questions people ask.

Is Roo Code still maintained?

No. The Roo Code documentation says the extension was shut down on 15 May, and its GitHub repository is archived. The notice points users to Cline, where Roo Code originated, and to a community fork.

Is Kilo Code a fork of Cline?

Indirectly. Kilo Code began as a fork of Roo Code, which was itself a fork of Cline. Its VS Code extension has since been rebuilt on a core shared with the Kilo CLI, which its repository describes as a fork of OpenCode.

Can Cline use Claude models?

Yes, with an Anthropic API key, through AWS Bedrock or Google Vertex AI, or through gateways such as OpenRouter. Cline also documents a Claude Code provider that runs your installed Claude Code binary; check Anthropic’s terms on subscription credentials before relying on it.

What is the Cline equivalent of Claude Code plan mode?

Plan mode. Cline reads and discusses but cannot change files or run commands until you switch to Act mode, which keeps the conversation and does the work.

Start with one thing.

There is nothing to set up first. Write one line and you’ve started.