Cloud/Create Connector/Agent Skills
Agent Skills
The second path in Create Connector: give AI the knowledge to use your app. Define reusable knowledge and instructions that AI agents can discover and follow, in Markdown, on the SKILL.md standard.
Agent Skills is the second path inside Create Connector, and it is the one for teaching. The console's card says exactly what it is: "Give AI the knowledge to use your app. Define reusable knowledge and instructions that AI agents can discover and follow." Its two tags are the how: Progressive disclosure and the SKILL.md standard. Where Connector Deploy ships code, this path ships understanding: no endpoints, no build step, just structured instructions written in Markdown.
Deploy Connector
Teach your AI new skills in under 30 seconds.
Skill Definitions
Full Markdown supported: headings, lists, code blocks, etc.
Full Markdown supported: headings, lists, code blocks, etc.
The mockup is the real editor. Add and remove skill cards, edit the name, description and Markdown body, then hit Deploy to Edge and read your connection token with the config JSON ready to paste into Claude, Cursor, VS Code or any other client.
Teach AI how to think
The education panel puts the philosophy in one line: "Skills are structured instructions that AI agents discover progressively, only loading what's needed for the current task. Less noise, better reasoning." That is progressive disclosure in practice: an agent does not swallow your whole handbook, it reads the skill's name and description first and loads the body only when the task needs it. The wizard is fast by design: "Teach your AI new skills in under 30 seconds."
The skill editor
The form has two parts. A Server Name, which is the connector agents will discover. And the Skill Definitions, one card per skill, each with three fields:
- Skill Name, a friendly name like "Customer Onboarding", capped at 64 characters;
- Description, one short line saying what the skill teaches the agent, because this is what the agent reads to decide whether to load the skill at all;
- Skill Body (Markdown), the instructions themselves. Full Markdown is supported: headings, lists, code blocks and more, up to a 50,000 character body.
Add Skill appends another card, and the trash button removes one (the last remaining card stays). The editor enforces the minimum honestly: "At least one complete skill (name, description, and body) is required." An empty skill card is not a deployable connector.
The deploy
The action bar holds the familiar emerald Deploy to Edge button. Behind it the console walks four phases: indexing the skill definitions, building the progressive disclosure tree, sealing the V8 Isolate sandbox, and going live on the global edge. Then the success screen takes over: your connector is Live!, with its Connection Token shown once, the "Save this token now. It won't be shown again." warning, the note that authentication is HMAC-SHA256 with the plaintext never stored, and the mcpServers config JSON ready to copy or download.
Connecting a client
Below the config, the success screen offers one tab per client: Claude, Cursor, VS Code, Antigravity and Others. The Claude tab shows where the config file lives on your OS and tells you to merge the mcpServers block and restart; the other tabs hand you the connector URL to paste, and Others lists ChatGPT, Windsurf, Cline, Continue and Zed. The same connector, five ways in.
Why this path exists
Tools let agents act; skills let them act well. The procedure your best support person follows, the rules your billing team applies, the checklist that keeps deployments safe: those are knowledge, and knowledge written once here becomes something every agent can discover and follow. Your skills deploy with the same edge runtime, the same token security and the same governance as every other connector, and they appear in your Library like any other connector, ready for Playground testing.
What comes next
The remaining Create Connector paths have their own guides, one at a time: OpenAPI / Swagger (paste one spec URL and every endpoint becomes a tool) and Manual Setup (no spec? add endpoints by hand, step by step).