Built-in Action Sets
Built-in actions are organized into collapsible sections, each with its own enable/disable toggle.Core Actions
These are always loaded into every agent session (unless disabled). They cover fundamental operations:| Action | Category | Description |
|---|---|---|
Read | Filesystem | Read files and directories |
Edit | Filesystem | Make targeted edits using search and replace |
Write | Filesystem | Create new files or overwrite existing ones |
Bash | System | Execute shell commands in a terminal |
Glob | Search | Find files matching glob and wildcard patterns |
Grep | Search | Search file contents using regular expressions |
AskUserQuestion | Interaction | Ask the user a question and wait for a response |
Extended Actions (On-Demand)
Extended actions are deferred — they aren’t loaded at session start. Instead, the agent discovers them viaToolSearch when it needs them. This keeps the base tool set lean.
| Action | Category | Description |
|---|---|---|
WebSearch | Search | Search the web for real-time information |
WebFetch | Search | Fetch and read content from a URL |
NotebookEdit | Filesystem | Edit Jupyter notebook cells |
TodoWrite | Planning | Write and manage a structured todo list |
EnterPlanMode | Planning | Enter plan mode for designing solutions |
ExitPlanMode | Planning | Exit plan mode and return to execution |
EnterWorktree | System | Enter a git worktree for isolated work |
TaskOutput | System | Read output from a background task |
TaskStop | System | Stop a running background task |
CronCreate | Scheduling | Create a scheduled or recurring task |
CronList | Scheduling | List all scheduled tasks |
CronDelete | Scheduling | Delete a scheduled task |
RenderOutput | Views | Render a reusable View artifact with structured input data |
Browser Actions
Browser automation is split into two layers: Delegation layer — what the main agent calls:| Action | Description |
|---|---|
CreateBrowserAgent | Create a new browser and run a task on it |
BrowserAgent | Delegate a task to an existing browser agent |
BrowserAgents | Run multiple browser tasks in parallel |
| Action | Description |
|---|---|
BrowserScreenshot | Capture a screenshot of the page |
BrowserNavigate | Navigate to a URL |
BrowserClick | Click an element by CSS selector |
BrowserType | Type text into an input element |
BrowserEvaluate | Execute JavaScript in the browser |
BrowserGetText | Get visible text content of the page |
BrowserGetElements | List interactive elements with CSS selectors |
BrowserScroll | Scroll the page up or down |
BrowserWait | Wait for page loads or animations |
The main agent never calls browser action tools directly. It delegates via
CreateBrowserAgent or BrowserAgent, and a sub-agent handles the low-level browser interactions autonomously.Apps
If you’ve created Apps, they appear here as an additional action set. Each App is exposed as aRenderOutput call that the agent can invoke to display data. Views have their own per-item permission toggles.
Enabling and Disabling Sections
Each action set has a toggle switch in its header. Disabling a section sets all of its actions todeny, which means agents cannot use any of them. Re-enabling restores them to always_allow.
You can also control permissions at a more granular level — see Permissions.
First-Party Integrations
Integrations are pre-configured connections for popular services. They appear in their own section of the Action Library.Google Workspace
Connect your account
Click Connect Google. A popup opens the Google OAuth consent screen. Sign in and grant the requested scopes.
you@gmail.com) and you can disconnect or switch accounts at any time.
Token refresh is handled automatically. When an access token expires, OpenSwarm uses the stored refresh token to obtain a new one before the next agent session starts. If refresh fails, you’ll be prompted to reconnect.
Twitter / X (via xbird)
Provide credentials
Click Connect and enter your
auth_token and ct0 cookie values from x.com. To find these: open x.com in your browser, press F12, go to Application → Cookies → x.com, and copy the values.@yourhandle) is displayed after successful connection.