# Capabilities

The registry of tools and behaviors an agent can use — execution, sandboxes, storage, data, safety, and optimization — as composable units.

[← Back to home](/)

## What a capability is

A capability is a self-contained unit that extends an agent. A single capability can contribute tools the model can call, additions to the system prompt, and mount points the agent operates against. Instead of hard-coding what an agent can do, you attach capabilities and the agent gains exactly those behaviors.

This keeps the agent definition small. An agent is configuration — a prompt, an optional model override, and the set of capabilities it is allowed to use.

## Where capabilities come from

Capabilities are resolved from four origins, so the same attachment mechanism covers built-in behavior and your own extensions:

- **Built-in** — first-party capabilities shipped with the platform, named in `snake_case` (for example `web_fetch`, `bashkit_shell`, `session_file_system`).
- **MCP servers** — tools and context exposed over the Model Context Protocol, surfaced as capabilities.
- **Registry skills** — reusable instruction-and-tool packages discovered from the workspace.
- **Declarative** — capabilities defined as configuration rather than code.

## Categories

Every built-in capability declares a category. The main groups are:

- **Core** — session basics, current time, and message metadata.
- **Execution and Sandboxes** — `bashkit_shell` for in-process shell, plus isolated sandboxes via Daytona, E2B, Deno, and Docker. See [Sandboxes](/platform/sandboxes).
- **File Operations and Storage** — `session_file_system` for a per-session filesystem and `session_storage` for key/value and encrypted secrets.
- **Data and Knowledge** — session-scoped SQL databases and knowledge bases.
- **Orchestration** — sub-agents, schedules, and agent handoff.
- **Optimization** — tool search, budgeting, parallel tool calls, and extended context.
- **Safety** — guardrails, prompt-injection canaries, and tool-call repair.
- **MCP Servers** — external tools and context attached over the Model Context Protocol.

## Attachment layers

Capabilities are attached at three levels, each narrowing the one above it:

- **Harness** — defaults that every agent on the harness inherits.
- **Agent** — capabilities for a specific role, on top of the harness defaults.
- **Session** — capabilities scoped to a single conversation.

## Related

- [Sandboxes](/platform/sandboxes) — the isolated execution environments behind the sandbox capabilities
- [Integrations](/platform/integrations) — external services Everruns connects to
- [Capability reference](https://docs.everruns.com/capabilities/) — every built-in capability with its tools, parameters, and dependencies