Skills

Skills are reusable capabilities you attach to agents. Each skill is a folder — a SKILL.md (instructions + metadata) plus optional scripts and resources — following Anthropic’s Agent Skills format. An agent opts in to the skills it needs; at run time the model can read a skill’s instructions and run its scripts.

Skills live as files in your connected repo under skills/<name>/, right alongside agents/. They run locally in your environment (no Anthropic code-execution sandbox) and work with any model.

Installing skills

Open Skills in the sidebar. Several sources, each commits the skill folder to your repo:

Anthropic knowledge-work skills — browse role-specific Agent Skills from anthropics/knowledge-work-plugins (sales, support, finance, legal, data, …) by work area and install with one click. These give an agent domain expertise it can draw on; pair them with a matching agent library starter.
skills.sh — browse and search the open Agent Skills directory right in the app and install with one click. (You can also paste a slug or GitHub URL for a skill that isn’t in the directory.)
Upload — a custom skill bundled as a .zip (a folder containing SKILL.md).
Import from the Claude API — copy a skill your team created via the Claude Skills API into the repo. Needs an Anthropic API key in Settings → LLM Providers.

Operators and workspace admins can upload custom skill bundles through Skills → New skill → Upload a custom skill, without needing an admin to upload on their behalf. Skills are shared across the workspace; uploading a bundle with an existing skill name updates that skill’s files.

Installing from GitHub or skills.sh, importing from the Claude API, and removing skills still require a workspace admin. Viewers cannot upload, install, or remove skills. Successful installs, uploads, and removals are recorded with the acting team member in the audit log.

Skill owners

The first team member to successfully upload or install a skill is recorded as its owner. Uploading or installing updates preserves the existing owner. Removing a skill clears its ownership, so a new installation gets a new owner.

The Skills list includes an Owner column and an Owner filter. Choose a team member to see their skills, Unassigned to find skills without an owner, or All owners to reset the filter. The skill detail page also shows its owner. Workspace admins can change the owner of any skill (including unassigned skills) from its detail page. The current owner can also transfer their skill to another workspace member, even if they are not an admin. Other members cannot reassign it. An admin can assign themselves; a non-admin owner must choose someone else. Once transferred, the previous owner can no longer reassign the skill unless they are an admin. Ownership changes are audited.

Skills remain shared across the workspace. Ownership does not grant upload, install, or removal permissions; the permissions above still apply. An owner with the viewer role can transfer ownership, but cannot modify skill content.

Skills installed before ownership tracking was introduced, or added directly to the repository, show Unassigned until an admin assigns an owner or their next successful upload or installation through the app. Deleting an owner’s user account also leaves their skills unassigned.

Opting an agent in

Add the skill’s folder name to the agent’s skills: field (the Tembo coding agent does this when you ask it to use a skill):

name: deck-builder
model: anthropic:claude-opus-4-7
instructions: |
  Build slide decks from the user's outline.
skills:
  - pptx
  - brand-guidelines

The agent page shows which skills an agent uses. A run resolves the named folders from the repo and mounts them; a declared-but-missing skill fails the run (same as a missing tools_module).

At run time

Skills are mounted with pydantic-ai-skills: the model gets tools to load a skill’s instructions, read its resources, and run its scripts. Skill scripts execute in the agent’s runtime — the same trust model as sidecar Python tools (your own, review-gated repo code). Nothing runs in an Anthropic-hosted container.