Coding agents and ACP editors (ACP)

wickrunAI speaks ACP (Agent Client Protocol) in both directions. Both need wickrunAI desktop 6.1 or later:

• Use OpenCode, Gemini CLI and other ACP agents as execution clients: pick one for Work and it acts on your computer, with your Work profile deciding what it may do and what it asks first.
• Use wickrunAI as the agent in Zed, JetBrains IDEs and Neovim: wickrunAI answers in the editor, with your routes, your Work profile and the confidentiality layer behind it.

Both are set up in Settings → Tools → Execution clients.

ACP agents as execution clients

Add one

1. Install the agent on your computer first, for example OpenCode or Gemini CLI.
2. Open Settings → Tools → Execution clients:
   • click "Discover" to add OpenCode (opencode acp) and Gemini CLI (gemini --acp) when they are installed;
   • or click "Add an ACP agent" and enter a name, an icon, the command and its arguments (for example opencode acp, or the program's full path with its arguments), an optional default working folder, and the names of environment variables to pass (for example OPENCODE_API_KEY; the values come from this computer and are not stored in settings).
3. Click "Check" to make sure it starts and to see its models.

The agent then appears in the on-device AI panel under the composer, in the model picker and in @ mentions.

What it is like

• Permissions: when it wants to edit files, run commands or go online, this computer's Work profile decides. Actions that need your OK show a card such as "Confirm OpenCode's action" with the folder, the files and the command. The agent's own "always allow" is never picked for you; standing permission lives only in your Work profile.
• Changes: the files it changed go into the run record, where you can review and roll them back in the code changes panel.
• Confidentiality layer: with masking on, it works in a masked copy; what it receives is masked, and results are restored before they are written back.
• Live documents: when the files are in a cloud workspace, it shows up in the live document as a member such as "R's OpenCode" while it works, and leaves a Work version when the run ends.

wickrunAI as the agent in your editor

Turn it on

1. In Settings → Tools → Execution clients, turn on "Let editors use wickrunAI as an agent (wickrun acp)".
2. The settings page shows Zed and JetBrains snippets for this computer, ready to copy. In the examples below, use the values from the settings page for the parts in angle brackets.
3. To use Work from the editor, the project folder must be authorized in Settings → Tools → Working directories; outside an authorized folder the session is chat only.

Zed

Add to ~/.config/zed/settings.json:

{
  "agent_servers": {
    "wickrunAI": {
      "type": "custom",
      "command": "<program path from the settings page>",
      "args": ["<launcher path from the settings page>", "acp"],
      "env": { "ELECTRON_RUN_AS_NODE": "1" }
    }
  }
}

Then run agent: new thread in Zed and pick wickrunAI.

JetBrains IDEs (AI Assistant)

Open the ACP agent configuration (acp.json) from the agent selector in AI Chat and add:

{
  "agent_servers": {
    "wickrunAI": {
      "command": "<program path from the settings page>",
      "args": ["<launcher path from the settings page>", "acp"],
      "env": { "ELECTRON_RUN_AS_NODE": "1" }
    }
  }
}

Neovim

avante.nvim:

acp_providers = {
  ["wickrunai"] = { command = "<wickrun path from the settings page>", args = { "acp" } },
}

CodeCompanion's ACP adapters take the same command and arguments.

If you have added the wickrun command to your PATH (see the VS Code extension guide), any editor can use the command wickrun with the argument acp.

What it is like

• Every session in the editor is a wickrunAI conversation titled "<editor> · <folder>", marked with its source, such as "From Zed", in the sidebar. Continue it in the editor or back in wickrunAI; it syncs like any other conversation.
• Files a turn changed show up in the editor and become that answer's work outputs, which you can review and roll back in wickrunAI's code changes panel.
• A turn you cancel in the editor stays as "cancelled" and is left out of later context.

Limits you may run into

• The wickrunAI desktop app must be running. When it is not, the editor gets a readable message.
• wickrunAI can only gate what an agent asks permission for. An agent that edits files or runs commands without asking is not stopped; its changes inside the folder are still recorded and can be rolled back. Commands do not run in a sandbox.
• With "review every change" on, ACP agents cannot run Work.
• ACP agents are not yet used for delegation, cloud workspace tasks or "use on-device AI" from the web.
• Editor sessions use your default Work profile and are not put into a project automatically.
• MCP servers the editor passes into a session are ignored.
• A file you edit in the editor and have not saved during a turn is overwritten by what wickrunAI writes back.
• A cancelled turn does not record its file changes, but files it already wrote stay on disk.