CrispyVibe

AI Agents

CrispyVibe connects AI coding agents directly inside your workspace using the Agent Conversation Protocol (ACP). You can chat with agents, approve their actions, and watch them work — all without leaving the IDE.

Supported agents

CrispyVibe supports three types of agents:

  • ACP-compatible agents — any agent that implements the Agent Conversation Protocol (including Kiro CLI)
  • Claude Code — direct integration with Anthropic’s coding agent
  • Codex — direct integration with OpenAI’s coding agent

Prerequisites

Install the agent’s CLI tool and make sure it is available on your system PATH. CrispyVibe discovers agents by looking for their executables at launch.

💡 Tip: If you install an agent CLI after launching CrispyVibe, restart the app so it can discover the new executable.

Connecting to an agent

  1. Open the ACP pane from the toolbar or the command palette.
  2. Select an agent from the dropdown.
  3. Select the project the agent should work in.
  4. Click Connect.

Agents that appear grayed out were not found on your PATH.

Auto-connect

Standalone ACP panes remember their configuration. When you reopen a VibeSpace, the pane reconnects to the same agent automatically.

Trust modes

Trust modes control whether the agent can act without your approval.

ModeBehavior
StandardEvery action requires your approval via a permission card
Full TrustAll actions are auto-approved

Escalating trust mid-session

When a permission card appears in Standard mode, click Allow Always to escalate to Full Trust for the remainder of the session.

Conversation

Type your message in the compose bar and press Enter to send. Click Cancel to stop the agent mid-response.

Response types

Agents can respond with:

  • Text — rendered as Markdown
  • Thoughts — internal reasoning shown in a collapsible block
  • Tool calls — grouped into expandable cards showing what the agent did
  • Plans — step-by-step outlines of intended work
  • Diffs — file changes shown in a diff viewer

Resend a message

Hover over a sent message and click Resend. This removes all subsequent messages and re-sends from that point.

Slash commands

Type / in the compose bar to see available slash commands for the connected agent.

Permissions

In Standard trust mode, the agent requests permission before taking actions. Each request appears as a permission card with three options:

  • Allow — approve this single action
  • Deny — reject this action
  • Allow Always — approve this action and escalate to Full Trust

Models and reasoning

Model selection

ACP-compatible agents negotiate their model during the connection handshake.

For direct integrations, you choose the model explicitly:

Claude Code models:

  • Sonnet 4.5
  • Opus 4.6
  • Sonnet 4
  • Haiku 4

Codex models:

  • GPT-5.4
  • GPT-5.3 Codex
  • GPT-5.3 Codex Spark
  • GPT-5.2 Codex
  • GPT-5.2

Reasoning levels

Control how much the agent reasons before responding:

  • Low — fastest responses, minimal reasoning
  • Medium — balanced (default)
  • High — more thorough reasoning
  • Max — maximum reasoning depth

Standalone pane vs Board tile

You can use an agent in two layouts:

  • Standalone pane — opens in the content viewer for focused, full-width work
  • Board tile — appears as a tile in the terminal board for monitoring alongside terminals

Settings

Global defaults

Open Settings → Agents to configure:

  • Default agent
  • Default trust mode
  • Default model and reasoning level
  • Custom agents (title, executable, and arguments)

Per-project overrides

Override agent defaults for a specific project in VibeSpace Settings.

Custom agents

Add custom ACP-compatible agents by specifying:

  • Title — display name in the agent dropdown
  • Executable — path or command name
  • Arguments — CLI arguments passed on launch

Troubleshooting

Agent not found in dropdown

  • Confirm the agent CLI is installed and available on your PATH.
  • Restart CrispyVibe to re-scan available executables.

Connection failed

  • Update the agent CLI to the latest version.
  • Open Developer Tools to check for connection errors.

Permission denied

The agent tried to access a file outside the project boundary. Move the file into the project or adjust the project root.

Agent stops responding

  1. Click Cancel to abort the current request.
  2. Click Reconnect in the pane header.

Known limitations

  • Conversations are not persisted within a single session by default. See Conversation History for cross-session persistence.
  • Multi-agent sessions are not supported — each pane connects to one agent.
  • Real-time collaboration between multiple users and agents is not available.