Agent Skills

Agent Skills are small Markdown files that teach an AI agent how to use a tool. They sit on disk next to the agent (e.g. ~/.claude/skills/), describe when to use a tool and how, and are loaded on demand when their trigger conditions match the task at hand.

Tektona ships two skills:

• tektona-cli installs, authenticates, manages sandboxes, SSH and VNC into them, mints preview URLs, and configures egress network policy from the shell.
• tektonactl drives the in-sandbox control tool: capture screenshots, click, type, manage the clipboard, and run named long-running PTY sessions inside a sandbox.

The two cross-reference each other, so an agent only loads the surface relevant to the task: sandbox lifecycle from outside, desktop or PTY automation from inside.

Prerequisites

You need the tektona CLI on your PATH before installing the skills. See the CLI overview for install options. In short:

npm install -g @tektona/cli

pnpm add -g @tektona/cli

bun add -g @tektona/cli

Then authenticate:

tektona api-key set tektona_xxx
tektona ctx set <org> <project>

Install the skills

Skills are distributed via GitHub. Use the Vercel skills CLI, with no install needed, run it through npx:

# Both skills
npx skills add tektona-ai/agent-skills

# Just one
npx skills add tektona-ai/agent-skills --skill tektona-cli
npx skills add tektona-ai/agent-skills --skill tektonactl

The CLI clones the tektona-ai/agent-skills repository (shallow), discovers each skill's SKILL.md, and symlinks them into your active agent's skill directory. For Claude Code that means:

~/.claude/skills/tektona-cli/SKILL.md
~/.claude/skills/tektonactl/SKILL.md

Other Agent Skills-compatible runtimes (Codex, GitHub Copilot CLI, Gemini CLI, …) are detected automatically. The skills themselves are runtime-agnostic, plain Markdown plus YAML frontmatter.

Skills evolve alongside the CLI. The published skills always match the currently supported CLI surface. Run npx skills update (see below) to pull in the latest after a CLI upgrade.

What the agent gains

Once the skills are installed, your agent will:

• Recognise that "spin up a sandbox and run my test suite" maps to tektona sandbox create -i ghcr.io/tektona-ai/desktop-x11:0.4.3 --egress-policy tektona/open … and follow up with tektona ssh to execute commands.
• Know that new sandboxes use tektona/dev unless overridden by project or org settings, and propose --egress-policy tektona/open or another allowed policy when broader egress is needed.
• Understand the difference between a token-bearing preview URL and a durable canonical URL on a --public sandbox, and pick the right flag for the situation.
• Inside a sandbox, drive the desktop with tektonactl desktop click/type/screenshot and manage long-running servers with tektonactl pty create/list/logs/kill.
• Recommend the official desktop image (ghcr.io/tektona-ai/desktop-x11) for VNC and computer-use workflows, pinned to a specific tag (the API rejects floating :latest references).

Update

Re-run the install command, or use the skills CLI's update flow:

npx skills update

This compares the installed skills against the latest published version and pulls in changes automatically.

Uninstall

Skills are symlinks. Remove them with the skills CLI:

npx skills remove tektona-cli
npx skills remove tektonactl

Or delete the symlinks directly from your agent's skill directory.

Triggering rules

Each skill declares when an agent should load it. Paraphrased from their YAML frontmatter:

Agents that respect the Agent Skills spec will only load a skill when its description matches the active task, so installing both is cheap.

What's next