# Agent Profiles & Variants

Agent profiles let you define multiple named variants for each supported coding agent. Variants capture small configuration differences (e.g., planning vs. coding, model choice, sandbox/approval levels) you can quickly select when creating an attempt or sending a follow‑up.

Profiles are used anywhere an agent runs:

- Onboarding: pick your default agent and optional variant
- Settings → General: set the default agent configuration
- Create Attempt dialog: choose agent and variant per attempt
- Follow‑ups: continue using the attempt's selected profile/variant

## Where To Configure

Open Settings → Agents. You can edit agent configurations in two ways:

- Form editor: guided fields per agent

![coding Agent Configurations](/images/coding-agent-configurations.png)

- JSON editor: edit the underlying `profiles.json` directly

![coding Agent Configurations JSON Editor](/images/coding-agent-configurations-json.png)

The page shows the exact file path where your config is stored (profiles.json). The app saves only your overrides; built‑in defaults remain available and cannot be deleted.

## JSON Structure

The configuration file is a JSON object with an `executors` map. Each executor (agent) contains one or more variants. Every variant holds a single object for that agent with its options.

Example (abridged):

```json
{
  "executors": {
    "CLAUDE_CODE": {
      "DEFAULT": { "CLAUDE_CODE": { "dangerously_skip_permissions": true } },
      "PLAN":    { "CLAUDE_CODE": { "plan": true } },
      "ROUTER":  { "CLAUDE_CODE": { "claude_code_router": true, "dangerously_skip_permissions": true } }
    },
    "GEMINI": {
      "DEFAULT": { "GEMINI": { "model": "default", "yolo": true } },
      "FLASH":   { "GEMINI": { "model": "flash",   "yolo": true } }
    },
    "CODEX": {
      "DEFAULT": { "CODEX": { "sandbox": "danger-full-access" } },
      "HIGH":    { "CODEX": { "sandbox": "danger-full-access", "model_reasoning_effort": "high" } }
    }
  }
}
```

Notes:

- Variant names are case‑insensitive and normalized to SCREAMING_SNAKE_CASE.
- `DEFAULT` is reserved and always present; you can add more variants (e.g., `PLAN`, `FLASH`).
- You can't remove built‑in executors or their built‑in variants, but you can override their values or add new variants.

## Per‑Agent Options (quick reference)

- CLAUDE_CODE
  - `plan: boolean` – enable planning mode
  - `claude_code_router: boolean` – route across multiple Claude Code instances
  - `dangerously_skip_permissions: boolean` – skip permission prompts (unsafe)
  - Docs: https://docs.anthropic.com/en/docs/claude-code/cli-reference#cli-flags
- GEMINI
  - `model: "default" | "flash"`
  - `yolo: boolean` – run without confirmations
  - Docs: https://google-gemini.github.io/gemini-cli/
- AMP
  - `dangerously_allow_all: boolean` – allow all actions (unsafe)
  - Docs: https://ampcode.com/manual#cli
- CODEX
  - `sandbox: "read-only" | "workspace-write" | "danger-full-access"`
  - `approval: "untrusted" | "on-failure" | "on-request" | "never"`
  - `oss: boolean`, `model: string`, `model_reasoning_effort: "low" | "medium" | "high"`, `model_reasoning_summary: "auto" | "concise" | "detailed" | "none"`
  - Docs: https://github.com/openai/codex
- CURSOR
  - `force: boolean`, `model: string`
  - Docs: https://docs.cursor.com/en/cli/reference/parameters
- OPENCODE
  - `model: string`, `agent: string`
  - Docs: https://opencode.ai/docs/cli/#flags-1
- QWEN_CODE
  - `yolo: boolean`
  - Docs: https://qwenlm.github.io/qwen-code-docs/en/cli/index

Common options supported by several agents:

- `append_prompt: string | null` – text appended to the system/prompt
- `base_command_override: string | null` – override the underlying CLI command
- `additional_params: string[] | null` – extra CLI args

Use these carefully; "dangerously_…" flags bypass confirmations and can perform destructive actions.

## Selecting Defaults vs. Per‑Attempt

- Default: Set under Settings → General → "Default Agent Configuration."
- Per Attempt: Use the create attempt dialog to pick a different agent/variant for that run.

## MCP Servers

MCP (Model Context Protocol) servers are configured per agent under Settings → MCP Servers. This is separate from profiles but complements them. See the [MCP Support](/docs/user-guide/vibe-kanban-mcp-server) guide.