Skills

The dashboard has a top-level Skills page with two tabs: Installed (the skills currently loaded in the workspace) and Marketplace (a catalog you can install from). That page is where everything lives — there is no separate page under Settings.

1. From the marketplace. Open Skills → Marketplace, browse the catalog, and click Install on the skill you want. The skill appears in the Installed tab immediately.
2. By hand. If you are writing your own skill, drop a folder with a SKILL.mdfile into the bot's skills/ directory. The bot does not watch the filesystem, so click Reload on the skill in the Installed tab (or POST /api/skills/<name>/reload) to make it visible. A bot restart also works.

To check which skills the bot currently sees, type /skills in any chat — it prints a text list of installed skills, nothing more. It is not an activation command: there is no pick step. To actually use a skill, describe what you want in plain language and let the bot switch itself in.

When the bot transitions into a skill, three things happen in order: it loads the instruction body into the system prompt, applies the tier override (if the skill declares one), and starts the MCP tool server (if the skill declares one). When it transitions out, the instructions drop from the next prompt and the MCP server shuts down after its idle timeout.

You do not drive any of this by hand — the bot does it through the skill_transition tool. But it helps to know the order: if something goes wrong (MCP server fails to start, tier did not change) you know which piece to look at.

What “active” means across turns

The resolver keeps the currently active skill sticky by writing its name to session metadata. On the next turn the bot sees the same skill is still active and does not have to reload it. The skill stops being sticky when the bot transitions to a different skill, the session is reset, or the skill is removed from the catalog (in which case the persisted name is cleared).

If you are writing your own skill, the instructions live in a file called SKILL.md. It has two parts: a YAML header with metadata and a Markdown body with the actual instructions. The body is what gets injected into the bot's prompt when the skill is active.

---
name: github-assistant
description: Work with GitHub via MCP
model_tier: coding
mcp:
  command: npx -y @modelcontextprotocol/server-github
  env:
    GITHUB_PERSONAL_ACCESS_TOKEN: ${GITHUB_TOKEN}
  startup_timeout: 30
  idle_timeout: 10
---

Use the available MCP tools to work with GitHub.
Focus on the repositories the user asks about.
When creating issues or PRs, summarize first and ask for confirmation.

The $${GITHUB_TOKEN} placeholder in the env block is resolved at load time — first from a vars: section in the same frontmatter, then from the process environment. The full recipe (token setup included) lives in Cookbook → Connect GitHub via MCP.