Tools Overview
The tools available to ArgentOS agents — command execution, browsing, memory, tasks, and more.
What Are Tools?
Tools extend your agent beyond text generation. When the agent needs to take an action -- run a command, search memory, check the weather, manage tasks -- it uses a tool. Tools are defined as JSON schemas and executed by the agent runtime.
Built-in Tools
ArgentOS ships with a core set of tools:
| Tool | Description |
|---|---|
exec | Execute bash commands and scripts |
browser | Automate web browsing (navigate, click, extract) |
memory_store | Save information to persistent memory (MemU) |
memory_recall | Search and retrieve from memory |
tasks_list | List tasks from the task queue |
tasks_add | Create a new task |
tasks_start | Mark a task as in progress |
tasks_complete | Complete a task |
tasks_block | Mark a task as blocked |
doc_panel | Open documents in the dashboard slide-out panel |
canvas | Send content to external node device screens |
sag | Speech audio generation (text-to-speech) |
How Tools Work
- The LLM returns a
tool_useblock specifying the tool name and input parameters - The agent runtime validates the input against the tool's JSON schema
- Tool policies are checked (sandbox mode, approval requirements)
- The tool handler executes and returns a result
- The result is sent back to the LLM as a
tool_resultblock - The LLM incorporates the result into its response
Tool Policies
Tools can be restricted based on the agent's security mode:
- Unrestricted: All tools available, no approval needed
- Sandboxed: Dangerous tools require user approval
- Locked: Only safe tools (memory, tasks) are available
See Sandboxing for details.
Extending Tools
Via Plugins
Plugins can register new tools. See Plugins.
Via Skills
Skills are loadable capability packages that bundle tools with prompts. See Skills.
Tool Naming Conventions
doc_panel= slide-out document panel in the web dashboard (for reports and documents)canvas= external node device screens (MacBook, external monitors)
These names are deliberately distinct to avoid confusion. The canvas-docs-enforcer plugin injects disambiguation instructions into the system prompt.