Capabilities

Capabilities are modular functionality units that extend Agent behavior. A Capability can contribute additions to the system prompt, provide tools, and modify execution behavior. Users can enable multiple capabilities on an Agent to compose functionality.

Overview

When you assign capabilities to an agent, those capabilities enhance what the agent can do:

System Prompt Additions: Some capabilities add instructions to the agent’s system prompt
Tools: Capabilities can provide tools that the agent can use during conversations
Behavior Modifications: Future capabilities may modify how the agent processes requests

Available Capabilities

Agent Instructions

Status: Available

Reads AGENTS.md from the session workspace and includes it in the system prompt. Content is re-read on every turn, so changes are picked up automatically.

No tools — this capability only injects context into the system prompt
File: /workspace/AGENTS.md (plain Markdown, max 32 KiB)
Use cases: Project conventions, coding style, build instructions, architecture notes
See Agent Instructions for full documentation

Current Time

Status: Available

Adds a tool to get the current date and time in various formats and timezones.

Tool: get_current_time
Formats: ISO 8601, Unix timestamp, human-readable
Use cases: Agents that need to know the current time or date

No-Op

Status: Available

A no-operation capability for testing and demonstration purposes. Does not add any functionality.

Use cases: Testing capability assignment workflow

Deep Research

Status: Coming Soon

Enables deep research capabilities with a scratchpad for notes, web search tools, and structured thinking.

System Prompt: Adds research scratchpad instructions
Use cases: Research-focused agents

Sandboxed Execution

Status: Coming Soon

Enables sandboxed code execution environment for running code safely.

System Prompt: Adds code execution instructions
Use cases: Agents that need to run code

File System Access

Status: Coming Soon

Adds tools to access and manipulate files - read, write, grep, and more.

System Prompt: Adds file system access instructions
Use cases: Agents that need to work with files

Web Fetch

Status: Available

Enables fetching content from URLs and converting HTML to markdown or plain text.

Tool: web_fetch
Features:
- Fetch web content with configurable timeouts (1s for first byte, 30s for body)
- Convert HTML to markdown or plain text
- Extract metadata (size, filename, last modified)
- Returns metadata for binary content (images, PDFs) instead of failing
Use cases: Agents that need to retrieve information from the web

Agent Skills (Built-in Discovery)

Status: Available

Enables agents to discover and activate skills from the session filesystem. Skills are instruction packages (SKILL.md files) uploaded to /.agents/skills/{name}/SKILL.md in the session VFS.

Tools:
- list_skills - Scan /.agents/skills/ for available skills
- activate_skill - Load a skill’s full instructions by name
Dependencies: Automatically includes File System Access
Features:
- Progressive disclosure (names first, full instructions on activation)
- Path traversal protection on skill names
- Bundled file listing for skills with extra assets
- Invalid SKILL.md files reported but don’t block discovery
Use cases: Project-specific skills uploaded to the session workspace, per-session skill discovery

See Agent Skills for workspace-based skills and Skills Registry for the API.

Docker Container (Experimental)

Status: Available (Development only, integration plugin)

Provides tools to run commands and manage files in a Docker container tied to the session. Registered as an external integration plugin via the inventory crate (integrations/docker/).

Tools:
- docker_exec - Execute shell commands in the container
- docker_read_file - Read files from the container filesystem
- docker_write_file - Write files to the container filesystem
- docker_logs - Get logs from the container
- docker_stop - Stop and remove the container
Features:
- Container is lazily started on first tool use
- Persists for session duration
- Session-isolated containers
- Configurable Docker image and working directory
Configuration:
- image: Docker image to use (default: mcr.microsoft.com/devcontainers/python:3.11)
- working_dir: Working directory inside container (default: /workspace)
Use cases: Agents that need to execute code or manage files in an isolated environment

Managing Capabilities

Via UI

Navigate to the Agent detail page
Find the Capabilities section in the sidebar
Enable or disable capabilities using the checkboxes
Reorder capabilities using the up/down arrows
Click Save to apply changes

The order of capabilities matters - capabilities are applied in the order shown, with earlier capabilities’ system prompt additions appearing first.

Via API

Create agent with capabilities:

curl -X POST http://localhost:9300/api/v1/agents \
  -H "Content-Type: application/json" \
  -d '{
    "name": "My Agent",
    "system_prompt": "You are a helpful assistant.",
    "capabilities": ["current_time", "web_fetch"]
  }'

Update agent capabilities:

curl -X PATCH http://localhost:9300/api/v1/agents/{agent_id} \
  -H "Content-Type: application/json" \
  -d '{
    "capabilities": ["current_time", "session_file_system"]
  }'

Get agent (includes capabilities):

curl -X GET http://localhost:9300/api/v1/agents/{agent_id}

List all available capabilities:

curl -X GET http://localhost:9300/api/v1/capabilities

Capability Application Flow

When a session runs, capabilities are applied as follows:

The agent’s assigned capabilities are fetched (ordered by position)
Each capability’s system prompt addition is collected
Each capability’s tools are collected
The final system prompt = capability additions + agent’s base prompt
All capability tools are made available to the agent

Best Practices

Order Matters: Place more important capabilities first - their instructions appear earlier in the system prompt
Minimal Capabilities: Only enable capabilities the agent actually needs
Test Combinations: Some capability combinations may produce unexpected behaviors
Check Status: Coming Soon capabilities cannot be enabled yet