Settings

The VS Code extension can be configured through the Settings window, opened by pressing the gear icon in Kilo Code. Changes apply across extension surfaces, including the sidebar and Agent Manager. The CLI can also use the same JSONC config files when you use it directly.

Configuring with the Agent

The fastest way to change your Kilo configuration is to ask the agent to do it for you. The agent has a built-in skill that understands the full kilo.jsonc schema and can read, create, and update your config files directly.

Examples of things you can ask:

• "Switch my default model to Claude Sonnet"
• "Disable the OpenAI and Groq providers"
• "Set up an MCP server for Figma"
• "Auto-approve all read and glob operations"
• "Create a custom agent for code review"

The agent will edit the appropriate config file (global or project-level) and explain what it changed. This works in both the CLI and VS Code extension.

Managing Settings

Kilo reads JSONC config from a global location (~/.config/kilo/kilo.jsonc) and from your project (kilo.jsonc, or .kilo/kilo.jsonc). All clients — CLI, VS Code, and JetBrains — read the same files.

If kilo.json or the legacy opencode.json, opencode.jsonc, or config.json files exist in the same locations, Kilo reads and deep-merges them as well. Clearing a setting in the Settings UI (for example, setting a model back to "Not set") removes it from every config file that contains it.

{
  "experimental": {
    "speech_to_text_model": "openai/whisper-large-v3-turbo"
  }
}

{
  "experimental": {
    "speech_to_text_base_url": "https://api.openai.com/v1",
    "speech_to_text_api_key": "sk-...",
    "speech_to_text_model": "whisper-1"
  }
}

{
  "hide_prompt_training_models": true
}

{
  "agent": {
    "compaction": {
      "model": "anthropic/claude-haiku-4-5"
    }
  }
}

{
  "reasoning_display": "preview"
}

{
  "terminal_command_display": "collapsed"
}

Sandbox

On macOS and Linux, the VS Code extension includes a dedicated Sandboxing settings tab. The sandbox is disabled by default. When enabled, it limits agent filesystem writes and can block outbound network access from model-originated tools. Windows users do not see these settings because Windows sandboxing is not supported.

See Sandboxing for setup instructions, the exact filesystem and network boundaries, and platform limitations.

Kilo Swarm

Kilo Swarm lets a main session and its task descendants, including nested subagents, exchange messages on a shared board. It uses the existing Task tool, not a separate agent runtime. The board is not shared with unrelated sessions, even in the same repository or worktree.

Kilo Swarm is on by default. Turn it off in the VS Code Agent Behaviour settings, or set shared_agent_board to false in kilo.jsonc.

Use it when agents can benefit from discoveries during work:

• Search races: agents try independent approaches to the same problem and share useful findings.
• Complementary teams: agents work on different parts of a feature and share constraints or results.

Straightforward tasks can stay solo. Enabling the board does not mean agents are always running or that every task needs a team.

Post message (board_post) stores a message on the shared board. Read messages (board_read) retrieves messages from the board explicitly. Activity notices are best-effort: a stored message does not prove that a recipient was notified, read it, or acted on it. Posting does not start or resume an agent, and normal task completion still returns results to the parent.

All participants can read the board history, including messages addressed to others. Recipient selection is not a privacy boundary. Peer messages do not grant user approval or change permissions; HOLD and VETO are advisory, not controls that pause or cancel work.

When a main session has board messages, open the Board icon in its task header to read them, refresh them, or reset the board. Only the owning top-level session can view or reset its board; child sessions and cloud sessions cannot. Reset clears visible messages only and does not stop agents or clear conversations. See Kilo Swarm communication for the board dialog, ownership rules, and recipient-state warnings.

Experimental Features

{
  "experimental": {
    "batch_tool": false,
    "openTelemetry": true,
    "disable_paste_summary": false,
    "mcp_timeout": 30000
  }
}

Task subagent model selection

Enable Task Subagent Model Selection in Settings → Experimental, or set experimental.task_model_selection to true in kilo.jsonc. It is off by default.

This lets you explicitly request a different model, provider, or reasoning effort for an individual subagent task. The agent keeps normal defaults unless you request an override; it does not select models autonomously for cost or complexity. See Per-task model selection.

Claude Code migration

Enable Claude Code Migration in Settings → Experimental to import supported global Claude Code configuration on the next backend start. It is off by default and runs once, with no automatic retry.

The migration imports:

• Global instructions from ~/.claude/CLAUDE.md into Kilo's global AGENTS.md.
• Standalone skills from ~/.claude/skills/ that contain only a SKILL.md.
• Top-level MCP server definitions from ~/.claude.json, disabled until you enable them.

Existing Kilo content takes precedence; conflicts and unsupported items are skipped. Your original Claude files are not changed or deleted. After the attempt, Kilo stops loading global Claude instructions and skills as a fallback, but project-level compatibility such as a repository's CLAUDE.md is unaffected. A notification reports the outcome and points to a receipt with imported, skipped, and failed items.