Installs into .claude/skills of the current project.
Are you the author of Mcp Connectors?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/danmo-ai-mcp-connectors)
---
name: mcp-connectors
source: builtin
description: Use product connectors (MCP under the hood) and auth gateways without stuffing every action schema into context.
---
# Connectors
Danmo Work mounts enabled **connectors** into the Agent Loop. Each action is
exposed as a tool named `mcp_<connector>_<action>` (see system
`<mcp-tool-naming>`). Skills may cite short names; always call the full `mcp_*`
name from the tool list.
## When to use what
- Temporary / one-off HTTP → `http_request`
- Product SaaS actions (Notion, GitHub, Feishu docs) → connector actions
- Auth-heavy SaaS with many APIs → prefer a gateway connector (Composio /
OpenConnector) instead of many raw HTTP calls
- Chat ingress (WeChat / Feishu / WeCom / QQ) is **not** a connector — it only
delivers messages
## Gateway meta-tools (Composio-style)
If the mounted connector exposes search / connect / execute meta-actions:
1. Search for the right upstream action
2. Connect / wait for OAuth if the user is not linked yet (`ask_user` with the
authorize link)
3. Execute; prefer batch execute when available
4. Do **not** dump secrets or tokens into replies
## Auth
If a connector returns unauthorized / 401:
1. Ask the user to open **Connectors** in the UI and complete auth (headers or
OAuth token) for that connector.
2. Do not invent tokens or put secrets in `http_request` unless the user
explicitly pasted a short-lived token for a one-off call.
## Permissions
- Connector actions are `external` risk — expect approval unless permission mode
allows them; dangerous shell and external connector actions may still ask.
- Never put API keys into tool arguments when a secret header / OAuth token
already exists on the connector.