Skip to main content

Skills, Commands, and MCP on Windows

Three ways to extend what agents know and can do:

  • Skills: know-how an agent loads when it needs it.
  • Commands: your own / shortcuts that type a prompt for you.
  • MCP servers: other programs whose tools agents can call.

Skills​

Skills work as Skills describes on the Mac, with the differences below. They're in Settings → Skills.

Turn skills on. A skill is available to agents only when its switch is on. Osaurus comes with six, all off to start: Research Analyst, Creative Brainstormer, Study Tutor, Productivity Coach, Content Summarizer, and Debug Assistant. Skills you create or import are on.

Get more skills.

  • Create writes your own: a Name, a one-line Description (what the model sees when deciding to load it), and the Instructions.
  • Import → From file… takes a SKILL.md, a .json, or a .zip.
  • Import → From GitHub… reads a repository with a .claude-plugin/marketplace.json. Enter owner/repo, click Find, tick the skills you want, and click Import selected. Only the skills are imported.

Use a skill. Agents see the list of skills that are on and load one when a task calls for it; this needs the agent's Tools ability. To use one yourself, type / and its name in the message box, or click + and choose Command or skill… (with the box empty). The skill applies to that one message.

Limit an agent's skills. On the agent's Abilities tab, open Skills and choose Only the skills I pick.

Different from the Mac: six built-in skills, off to start; no tabs, export, or reference-file editing; agents can't edit skills themselves; Claude plugins aren't on Windows.

Commands​

A command is a prompt you use often, saved as a / shortcut. In Settings → Commands, click Create and give it a Name (which becomes /name), an optional Description and Icon, and the Template text.

Type / in the message box and pick it: its text goes into the box, and you carry on typing before you send. Nothing is sent until you do.

Osaurus also has /clear (new conversation), /model (switch the model), /folder (attach a workspace folder), and /help.

MCP servers​

An MCP server is a program that offers tools, such as access to a service or your files. Agents can call its tools like any others. They're in Settings → MCP servers.

Add a server​

Click Add server, then fill in:

  • Id: a short name, used as the start of each tool's name. It can't be changed later.
  • Name: shown in the list.
  • Transport: stdio for a program on your PC, or http for a server at a web address.

For stdio, enter the Command (such as npx), its Arguments, and any Environment variables, then choose where it runs:

  • This PC: with your Windows account's access to your files and apps. Only for a server you trust.
  • The sandbox: each agent that uses it gets its own copy, running in that agent's Linux home in the sandbox. Linux commands (npx, uvx, python3) work there; Windows paths don't. Its tools are available only while the sandbox is on.

For http, enter the URL. For a server that signs you in, set Authentication to OAuth: after saving, click Sign in on its row to finish in your browser. For a key or token, add it under Headers.

List the names of any secret environment variables or headers under Secret keys: they're stored encrypted and never shown again.

Click Save. The row shows Connected · N tools when it works.

Choose its tools​

Expand Tools on a server's row to turn single tools on or off. For an agent to call only some tools, open its Tools card, set Which tools to Manual, and list them.

Whether agents ask before calling a server's tools is set in Settings → Behavior → Tool approvals (Run automatically by default, or Ask each time), and each agent can override it on its Permissions tab.

A tool call that hears nothing from its server for 45 seconds stops; a server that reports progress can keep going for up to 10 minutes.

Use Osaurus's tools from another app​

The osaurus mcp command lets other MCP apps, such as an AI coding tool, call the tools of the servers you've added to Osaurus. See osaurus mcp.

Different from the Mac: no built-in directory of services; no Test Connection; timeouts aren't adjustable; each tool is on or off rather than having its own ask/auto policy; stdio servers can run in the sandbox.

Under the hood​

  • Skills are stored as skills\<name>\SKILL.md in the data folder. The system prompt lists the skills that are on; the model loads one with the skill_load tool.
  • MCP tools are named <server id>_<tool>. With automatic tool selection, their full descriptions aren't sent with every message: the model sees their names and loads a tool's details when it needs it, which saves tokens with large servers.
  • A sandboxed stdio server is one process per agent, running as that agent. Its environment variables reach it through a temporary file, never a command line.
  • The local server also exposes your servers' tools over HTTP at /mcp/tools and /mcp/call.