Agent Toolsets

What are Toolsets?¶

Toolsets in LLMling define groups of related tools that agents can access. Rather than configuring individual tool permissions, you enable entire categories of functionality through toolset configurations.

Think of toolsets as "skill packages" that give agents specific capabilities - from file operations to process management to agent coordination.

Available Toolsets¶

Agent Management (agent_management)¶

Enables agents to discover, coordinate with, and manage other agents:

agents:
  coordinator:
    toolsets:
      - type: agent_management

Provides tools:

• list_available_agents - Discover other agents in the pool
• list_available_teams - Discover available teams
• delegate_to - Assign tasks to other agents
• ask_agent - Ask other agents directly
• add_agent - Add new agents to the pool
• add_team - Create new teams
• connect_nodes - Connect agents/teams in workflows
• create_worker_agent - Create worker agents as tools
• spawn_delegate - Create temporary delegate agents

File Access (fsspec)¶

File system operations via fsspec (supports local, S3, GCS, etc.):

agents:
  reader:
    toolsets:
      - type: file_access  # Local filesystem by default, or use url: s3://bucket, etc.

Provides tools:

• read_file - Read files (text or binary)
• list_directory - List directory contents with filtering
• write_file - Write content to files
• delete_path - Delete files or directories
• edit_file - Edit files with smart matching
• download_file - Download files from URLs

Resource Access (resource_access)¶

Access to LLMling resources and configurations:

agents:
  assistant:
    toolsets:
      - type: resource_access

Provides tools:

• load_resource - Load resource content
• get_resources - Discover available resources

Code Execution (code_execution)¶

Execute Python code and system commands:

agents:
  developer:
    toolsets:
      - type: code_execution

Provides tools:

• execute_python - Execute Python code (WARNING: No sandbox)
• execute_command - Execute CLI commands

Process Management (process_management)¶

Start and manage background processes:

agents:
  build_manager:
    toolsets:
      - type: process_management

Provides tools:

• start_process - Start background processes
• get_process_output - Check process output
• wait_for_process - Wait for process completion
• kill_process - Terminate processes
• release_process - Clean up process resources
• list_processes - Show active processes

Tool Management (tool_management)¶

Register and manage tools dynamically:

agents:
  admin:
    toolsets:
      - type: tool_management

Provides tools:

• register_tool - Register importable functions as tools
• register_code_tool - Create tools from code

User Interaction (user_interaction)¶

Direct interaction with users:

agents:
  assistant:
    toolsets:
      - type: user_interaction

Provides tools:

• ask_user - Ask users clarifying questions

History (history)¶

Access conversation history and statistics:

agents:
  analyst:
    toolsets:
      - type: history

Provides tools:

• search_history - Search conversation history
• show_statistics - Display usage statistics

Integrations (integrations)¶

External service integrations:

agents:
  integrator:
    toolsets:
      - type: integrations

Provides tools:

• add_local_mcp_server - Add local MCP servers
• add_remote_mcp_server - Add remote MCP servers
• load_skill - Load Claude Code Skills

Common Patterns¶

Basic Assistant¶

agents:
  assistant:
    model: openai:gpt-4
    toolsets:
      - type: resource_access
      - type: file_access
      - type: user_interaction

Team Coordinator¶

agents:
  coordinator:
    model: openai:gpt-4
    toolsets:
      - type: agent_management
      - type: history
    system_prompts:
      - You coordinate tasks across multiple agents

Developer Agent¶

agents:
  developer:
    model: anthropic:claude-3-5-sonnet-20241022
    toolsets:
      - type: file_access
      - type: code_execution
      - type: process_management
      - type: tool_management
    system_prompts:
      - You are a software developer with full system access

Restricted Agent¶

agents:
  restricted:
    model: openai:gpt-4-mini
    toolsets: []  # No toolsets = only predefined tools
    tools:
      - calculator
      - web_search

Security Considerations¶

Toolsets provide different levels of system access:

Low Risk:

• fsspec - File system operations (configurable scope)
• resource_access - Access to configured resources
• user_interaction - User prompts only

Medium Risk:

• agent_management - Can create and coordinate agents
• history - Access to conversation data
• integrations - External service access

High Risk:

• code_execution - Can execute arbitrary code
• process_management - System process control
• tool_management - Can modify available tools

Always use the principle of least privilege - only enable toolsets that agents actually need.

Custom Toolsets¶

You can also create custom toolsets by implementing your own provider:

agents:
  specialized:
    toolsets:
      - type: custom
        import_path: "mypackage.toolsets.SpecializedTools"

See the Tools documentation for implementation details.