Every skill is served at a Streamable-HTTP MCP endpoint:
https://<your-domain>/mcp/<workspace>/<endpoint-name>?token=<token>Human-readable endpoint names
The endpoint name is the handle in the URL. Edit it on the skill page (lowercase letters, numbers, hyphens; unique in your workspace). Changing it moves the URL — the old one stops working immediately, so reconnect your agents. Connect snippets use this name as the server key (e.g. "rd-tax-credit": { "url": … }), so an agent’s config is self-describing.
Tokens & per-consumer keys
- The master endpoint token works for your own agents.
- Issue named, revocable per-consumer keys so each recipient gets their own metered key — the recipe never leaves the runtime. Revoke a key and that consumer stops instantly.
- Rotate the master token to invalidate the current URL.
Tools exposed to agents
The endpoint exposes a small, query-mediated tool set — agents discover and pull only what they need; the full skill (and any files) never leave:
list_sections— the manifest of the skill’s workflow sections.get_section— one section by title.query— the most relevant slice for a natural-language question.list_branches— the skill’s branches (e.g. Clients, Projects) and when to use each.list_nodes— the nodes inside a branch.get_node— one node’s content (e.g. a specific client’s overlay).
The skill’s workflow can point agents into the tree — e.g. “if a client overlay exists, call get_node before proceeding.”
Connecting an agent
See Connect your agent for ready-to-paste config for Claude Code, Claude Desktop, Cursor, Windsurf, Codex, the Claude/OpenAI Agent SDKs, STDIO via mcp-remote, and the MCP Inspector.