Skip to content
Agentic Coding Toolkit

Cursor

ACT on Cursor ships as a local plugin that exposes ACT skills.

Install

Section titled “Install”
Terminal window
cd ~/.agentic-coding-toolkit && ./scripts/install.sh --tool cursor

What Gets Installed

Section titled “What Gets Installed”
LocationIncludes
~/.cursor/plugins/local/Symlinked ACT local plugin directory
~/.config/agentic-coding-toolkit/Shared ACT settings and copied runtime helper

Verify

Section titled “Verify”

After installing or updating ACT for Cursor:

  1. Restart or reload Cursor
  2. Open Agent chat
  3. Type /act-

You should see ACT’s available act-* skills.

Skills

Section titled “Skills”

Invoke ACT skills from Agent chat using the same canonical names documented across the site, for example:

Terminal window
/act-create-issues ai_specs/0001-settings/spec.md
/act-implement ai_specs/0001-settings/work-items/01-settings-screen.md

Hooks

Section titled “Hooks”

ACT hooks are not currently supported on Cursor.

Permissions

Section titled “Permissions”

For the strategy comparison, see Permissions & Safety.

Strategy 1: Permission rules

Section titled “Strategy 1: Permission rules”

Cursor supports permission rules through permissions.json and Run Mode settings.

Use Allowlist when you want predictable terminal and MCP allowlists for ACT. Use Auto-review when you want Cursor’s classifier to decide whether matching tool calls can run, optionally guided by autoRun instructions.

permissions.json can be configured globally or per workspace:

  • ~/.cursor/permissions.json
  • <workspace>/.cursor/permissions.json

When both files exist, Cursor concatenates arrays from both files. When permissions.json defines an allowlist, it overrides the corresponding in-app allowlist for that allowlist type.

For ACT, the useful minimum is:

{
  "terminalAllowlist": [
    "git",
    "flutter",
    "dart",
    "node",
    "gh"
  ]
}

Cursor also supports mcpAllowlist for MCP tools:

{
  "mcpAllowlist": [
    "github:*",
    "linear:list_issues"
  ]
}

In Auto-review mode, Cursor also supports autoRun instructions for classifier guidance:

{
  "autoRun": {
    "allow_instructions": [
      "Allow read-only inspection commands for project files and build outputs."
    ],
    "block_instructions": [
      "Ask before deleting files, changing shell configuration, or touching credentials."
    ]
  }
}

autoRun steers Cursor’s classifier; it is not an enforcement boundary.

To make allowlists effective, enable a Cursor Run Mode such as Allowlist or Auto-review in Cursor Settings:

Cursor Settings showing Agents > Auto-Run with the Auto-Run Mode dropdown open

DSP / yolo equivalent

Section titled “DSP / yolo equivalent”

Cursor does not call its bypass mode DSP/yolo. Its closest equivalent is Run Everything, which skips Cursor’s safety checks. Treat it with the same isolation principle as DSP / yolo mode: prefer a non-sensitive account, remote dev machine, or containerized environment when appropriate.

Remote machines and containers

Section titled “Remote machines and containers”

You can run Cursor against a remote workspace or isolated environment where that fits your workflow. Treat the remote host or container as the isolation boundary. See Permissions & Safety for the full VPS and Docker tradeoffs.

This reduces repeated terminal approval prompts for ACT skills. Cursor’s built-in read/write/search tools are still governed by Cursor’s Run Mode rather than permissions.json.

Troubleshooting

Section titled “Troubleshooting”
  • permissions.json only takes effect when Cursor Run Mode is enabled.
  • If both user and workspace files exist, remember that arrays are concatenated.
  • If permissions.json defines an allowlist, the corresponding in-app allowlist is replaced by the file-defined entries.
  • Cursor CLI permissions are separate from Cursor editor permissions.

Relevant Cursor docs:

Section titled “Related”