Skip to main content
Every Skydive agent starts with a built-in toolset. It isn’t limited to that set. When a job needs something the agent doesn’t have, it gives itself the capability and keeps going. Otis connecting tools and MCP servers

​
Built-in tools

Out of the box, every agent can:

Run code & commands

Execute shell commands and scripts in its sandbox.

Read & write files

Create, read, and edit files: documents, data, code.

Browse the web

Drive a real browser to navigate, click, and extract.

Search the web

Find and fetch up-to-date information.

Host web apps

Run a live web server you can visit.

Process media

Work with images, audio, PDFs, and more: transcribe, convert, transform.

​
It can give itself new tools

The built-in set is the floor, not the ceiling. The agent has a real computer, so it can reach for whatever a task demands:
  • Install a command-line tool and use it.
  • Call any API, public or internal, once you’ve connected the service or given it a key.
  • Write and run its own code to compute, transform data, or automate a one-off job.
  • Add a new integration for itself when a task needs a system it isn’t connected to yet.
You don’t pick from a fixed menu of features. You describe the outcome, and the agent assembles the means.
No one set that up ahead of time. The agent reads the docs, writes the code, and uses the access you granted.

​
Connecting to more systems with MCP

One way an agent extends itself is MCP (Model Context Protocol): the open standard for connecting agents to external tools and data sources. If a capability exists as an MCP server, your agent can plug it in and use it.

​
What MCP adds

  • Tools that already exist as MCP servers (hosted or local)
  • Connections to internal systems and data sources
  • Provider-specific actions exposed through their own MCP servers

​
Authenticated MCP servers

Many MCP servers require authorization. You authorize once. A proxy then attaches the credential to every future request automatically. Tokens never sit in plain config and never appear in the agent’s prompts. See Connecting services. The Connect card for an MCP server grants access at whatever level that MCP app asks for (often the whole workspace or account). It is not the same as the OAuth connect flow some apps use, where you pick individual pages or records. On MCP Connect, you will not choose items one by one, even for apps (like Notion) that offer a picker on the regular OAuth path.

​
Choosing the right tool

The agent decides which tool fits each step. It searches when it needs facts. It browses when it needs to interact with a page. It runs code when it needs to compute or transform. It reaches for a new tool when nothing built-in fits. You describe the outcome; it picks the means.

Browser use

When the agent needs to act on a real web page.

Connecting services

Authorize the accounts and APIs your tools need.

MCP config in the agent repository

MCP connections and custom tools are files in the agent’s git repository. Inspect or edit them directly.

Platform CLI reference

Every command the agent can run inside its own sandbox.