Mistral Vibe

Mistral Vibe follows a strict hierarchy to resolve conflicting instructions. When instructions from different sources clash, the source with the lower number in this list takes precedence:

1. Critical instructions (Never overridable)
2. User messages (Recent messages override older ones)
3. Repo AGENTS.md files (Files on the path from the task up to the repo root; closer to the task wins)
4. The user's AGENTS.md
5. Overridable defaults (System prompt defaults)
6. Skills / MCP output
7. External data (Web, fetched content - treated as data, not instructions)

An instruction is considered active only if it is not overridden by a higher-priority source.

Vibe uses vibe.app_server as its central harness and runtime boundary. All delivery surfaces (Textual, ACP, or programmatic modes) act as clients to this server via a typed JSON-RPC 2.0 protocol.

Key architectural principles:

• Boundary: Delivery surfaces never interact with AgentLoop directly; they exchange serialized JSON values.
• Separation of Concerns: vibe.core handles model and tool execution, while vibe.app_server manages sessions, turns, callbacks, resources, and persistence.
• State Ownership: The server is authoritative for runtime construction, session identity, tools, permissions, and effective configuration. Clients are responsible for UI/UX elements like widgets, layout, and keyboard/prompt editing.

Mistral Vibe is architected with a separation between the core engine and various delivery surfaces. The vibe/core package contains the reusable engine implementations (agent loops, tools, LLM backends, etc.), while vibe/app_server acts as the runtime composition root.

Delivery surfaces include:

• vibe/cli: Terminal UX, Textual app, widgets, slash-commands, and voice UI.
• vibe/acp: Agent Client Protocol translation and presentation.
• vibe/setup: First-run and onboarding flows.
• Programmatic entry points: Direct consumption of the app-server client API.

To interact with the system, use the public facades provided by the app server rather than private modules.

To configure proxy settings for environments requiring network traffic to pass through a proxy server, use the interactive /proxy-setup command within the CLI. After completing the form, you must restart the CLI for the changes to take effect.

1. Type `/proxy-setup` and press Enter
2. Fill in the variables you need, then press **Enter** to save or **Escape** to cancel
3. **Restart the CLI** for changes to take effect

The protocol supports three message directions. When the server requires client participation (e.g., for a tool that requires local filesystem access), it uses the following pattern:

1. Server Request: The server sends a typed request via callback/call or an advertised clientTool/* operation.
2. Client Acknowledgment: The client acknowledges delivery of the request.
3. Client Response: The client sends the semantic answer back to the server using a callback/respond request. For client-tools, this response includes the result of the requested operation (e.g., terminal output).

Mistral Vibe requires the JetBrains AI Assistant extension.

For JetBrains version 2025.3 or later:

1. Navigate to Tools > AI Assistant > Agents in settings.
2. Search for Mistral Vibe and click install.
3. Select Mistral Vibe from the AI Assistant agent selector.

For legacy JetBrains versions (using acp.json): Add the vibe-acp command to your acp.json configuration under agent_servers and select the agent in the AI Chat selector.

{
  "agent_servers": {
    "Mistral Vibe": {
      "command": "vibe-acp"
    }
  }
}

Mistral Vibe includes several agent profiles tailored for different workflows. You can select an agent at runtime using the --agent flag or set a default in your config.toml.

Available Agents

• default: Standard agent; requires approval for tool executions. Best for general use.
• plan: Read-only agent for exploration and planning. Auto-approves safe tools like grep and read.
• accept-edits: Auto-approves file edits only (write_file, edit). Useful for refactoring.
• auto-approve: Auto-approves all tool executions. Use with caution.
• lean: Only available if listed in installed_agents.

Usage

To run with a specific agent:

vibe --agent plan

To set a default agent in config.toml:

default_agent = "plan"

Note: You can override any agent's safety settings by passing --auto-approve or --yolo during a session.

You can override default prompts by placing markdown files in ~/.vibe/prompts/ (or project-local .vibe/prompts/).

• System Prompts: Replace the default prompts/cli.md by setting system_prompt_id in config.toml to the filename (without .md).
• Compaction Prompts: Replace the default prompts/compact.md by setting compaction_prompt_id in config.toml to the filename (without .md).

Custom Instructions: You can also add project-specific instructions by creating AGENTS.md files in your project directory or ~/.vibe/AGENTS.md. Local files override global ones.

system_prompt_id = "my_custom_prompt"
compaction_prompt_id = "my_compaction_prompt"

The prepare_release.py script automates the release workflow by:

1. Building the release branch from the previous public release tag.
2. Cherry-picking commits from matching -private tags.
3. Squashing commits into a single release commit (default behavior).
4. Freezing the full transitive dependency graph into pyproject.toml using the current uv.lock.