SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsconnect-ai-manage-mcp

connect-ai-manage-mcp

Agents & models

View repository

Updated 10-16 08:34

About this skill

CData Connect AI — Manage the Platform over MCP

This skill administers connections and toolkits in CData Connect AI through the Management MCP server (https://mcp.cloud.cdata.com/mcp/mgmt). It is the MCP-native counterpart to connect-ai-manage: same job for the connection/toolkit slice, but driven by MCP tools instead of the raw /api/ui/* REST API. It does not query data, and it covers only what the Management MCP exposes — see the routing and scope tables below.

Because the Management MCP is OAuth-only for connection creation, real credentials never pass through the chat: authentication happens in the browser OAuth flow, and any non-OAuth source is handed off to the admin UI. Only non-secret configuration values (an instance URL, a domain, a subdomain) are ever collected in the conversation.

Where this skill sits (routing)

The user wants to… Use
Create an OAuth connection, test connections, or build/shape toolkits — over MCP this skill
Full platform admin: workspaces, jobs, users/roles, PATs, billing — or delete a connection/toolkit, or edit a custom tool's SQL connect-ai-manage (direct REST)
Create a non-OAuth connection (Basic / API token / etc.) Admin UI — this skill relays the source's newConnectionUrl
Read/write data in a source, with an MCP connector present connect-ai-base + the connector skill (connect-ai-<source>)
Read/write data with no MCP connector available connect-ai-direct
Set Connect AI up in Claude Code for the first time connect-ai-onboarding-claude-code

connect-ai-manage vs. connect-ai-manage-mcp. connect-ai-manage is the broader, standalone REST skill (Auth0/PAT) covering the whole admin surface — it is the only path for workspaces, jobs, users, PATs, billing, and for hard deletes. This skill is narrower and MCP-native: it does connections + toolkits through a supported, governed server, with OAuth-delegated auth. Prefer this skill when the Management MCP is connected and the task is connection/toolkit work; fall back to connect-ai-manage for anything outside that subset.

Ground rules

  1. Never act on activation. Loading this skill produces no banner, no menu, no tool call. Act only on a concrete connection/toolkit request, and confirm the server is reachable (Step 0) first.
  2. Everything goes through the Management MCP. This skill never calls a vendor API directly and never scrapes tokens; it only invokes the mcp/mgmt tools.
  3. OAuth-only connections. create_connection supports OAuth authentication only. Never pass an AuthScheme such as Basic, APIToken, PAT, or UserToken, and never ask the user to paste a password, API key, or client secret into the chat. For a non-OAuth source, relay the source's newConnectionUrl and stop.
  4. list_available_sources is the source of truth. Never tell a user a source is unsupported from prior knowledge — call list_available_sources first.
  5. No hard deletes here. The Management MCP has no delete tools. Deactivating (set_toolkit_active false, set_custom_tool_active false) is the closest — a deactivated toolkit returns 423 to MCP clients but still exists. For a permanent delete, or to edit a custom tool's SQL/parameters, send the user to the admin UI (connect-ai-manage covers connection/toolkit deletes over REST). Still confirm before deactivating.

Step 0 — Confirm the Management MCP is connected

Call tool_search for a Management-MCP tool name — search "list_toolkits" or "create_connection". These tools are unique to the mcp/mgmt server, so a hit confirms it is registered and authenticated.

  • Found → proceed to the operation.
  • Not found → the server isn't connected. Do not work around it (no raw HTTP to mcp/mgmt, no DevTools token). Tell the user the Management MCP isn't connected and offer to register it (Setup below). Data-MCP tools (getCatalogs, queryData) do not count — this skill needs the mgmt server specifically.

Setup — register the Management MCP (only if Step 0 found nothing)

The Management MCP uses the same authentication as the data MCP — a Personal Access Token over HTTP Basic (default), or browser OAuth. The only difference from onboarding is the URL: https://mcp.cloud.cdata.com/mcp/mgmt.

Reuse the registration flow from connect-ai-onboarding-claude-code verbatim, substituting the mgmt URL and a distinct server name. PAT path (credential stays in the user's own terminal, never in chat):

# The user creates a PAT in the Connect AI console (Settings → Access Tokens / Integrations → Claude Code),
# then runs BOTH lines in THEIR terminal, so the PAT stays local and never enters the chat.
CRED=$(printf '%s' 'you@example.com:YOUR_PAT' | base64 | tr -d '\n')
claude mcp add --scope user --transport http connect-mgmt https://mcp.cloud.cdata.com/mcp/mgmt --header "Authorization: Basic $CRED"

OAuth alternative (short-lived, self-refreshing) — register without the header, then authenticate:

claude mcp add --scope user --transport http connect-mgmt https://mcp.cloud.cdata.com/mcp/mgmt
# then in Claude Code:  /mcp → connect-mgmt → Authenticate

Then have the user restart Claude Code so the tools load, and re-invoke the request. See the onboarding skill for the full PAT-vs-OAuth walkthrough, callback-URL troubleshooting, and the "can't reach the CLI" fallback.

Connections

The connection tools, and the exact order to use them in:

Goal Tool Notes
List saved connections list_connections returns id, name, source type
List connectable sources list_available_sources authoritative; returns each source's newConnectionUrl
Inspect a source's full config schema get_source_properties(source) basic + advanced fields; for "what can I configure" questions
Get just the required fields to ask for prepare_connection(source) returns requiredFields (internal name, display name, description)
Create the connection create_connection(name, source, properties) OAuth-only; properties keyed by exact field name
Verify a connection works test_connection(id) run after create + after the user finishes OAuth sign-in

Create-a-connection workflow (OAuth sources)

  1. Resolve the source. Call list_available_sources and match the user's request to a real source internal name. If it isn't in the allow-list of OAuth sources, relay its newConnectionUrl verbatim and stop — non-OAuth setup happens in the admin UI.
  2. Gather the required fields. Call prepare_connection(source). For each requiredFields entry, ask the user in chat for the actual value. These are non-secret config fields (URL, domain, instance). Never invent values, and never accept or request a password/API key/secret here — if the source needs one, it isn't an OAuth source (step 1).
  3. Create. Call create_connection(name, source, properties) with every required field set. Do not send an AuthScheme. The tool validates and returns an actionable error if a field is missing.
  4. Sign in. The connection now needs its browser OAuth sign-in (done by the user, outside the chat). Point them to it.
  5. Test. Once they've signed in, call test_connection(id) and report success or the specific error returned.

Allow-listed OAuth sources (as of 2026-09): GitHub, GoogleCalendar, HubSpot, JIRA, SharePoint, Slack, Asana, ExactOnline, ExcelOnline, Facebook, FacebookAds, GoogleAds, GoogleAnalytics, GoogleDrive, GoogleSheets, LinkedInAds, Office365, OneNote, QuickBooksOnline, Salesforce, Square, SurveyMonkey, Xero, ZohoCRM. Always confirm against list_available_sources rather than this list.

Toolkits

A toolkit is a bundle of connections exposed to Data Copilot / MCP clients — and every toolkit is itself a remote MCP server (mcp_remote_server_url, returned by create_toolkit, get_toolkit, list_toolkits). So this skill can assemble a scoped MCP server that connect-ai-base then queries.

Goal Tool
List toolkits (incl. mcp_remote_server_url, isActive, tool counts) list_toolkits
Full state of one toolkit get_toolkit(toolkit_id)
Create an empty toolkit create_toolkit(name)
Rename rename_toolkit(toolkit_id, name)
Set server-level instructions / guardrails (empty string clears) update_toolkit_server_instructions(toolkit_id, server_instructions)
Activate / deactivate (deactivate = closest to delete) set_toolkit_active(toolkit_id, active)
List the connections in a toolkit list_toolkit_tools(toolkit_id)
Full op state for one connection in a toolkit get_toolkit_tool(toolkit_id, tool_id)
Attach a connection to a toolkit add_tool_to_toolkit(toolkit_id, connection_id)
Toggle one op set_tool_op_enabled(toolkit_id, tool_id, kind, op_name, enabled)
Toggle every op of a kind set_all_tool_ops_enabled(toolkit_id, tool_id, kind, enabled)
Set per-source-op AI instructions update_source_tool_instructions(toolkit_id, tool_id, source_op_name, instructions)
List custom SQL tools list_custom_tools(toolkit_id)
Activate / deactivate a custom tool set_custom_tool_active(toolkit_id, custom_tool_id, active)

Notes that change how you drive these:

  • kind is universal or source. Universal ops are the built-in SQL-flavored tools (execute_select, get_tables, …); source ops are driver-specific (JIRA's get_issue, …). Auto-included universal ops are unaffected by the bulk toggle.
  • tool_id is the connection's id inside the toolkit. add_tool_to_toolkit takes a connection_id (from list_connections), which becomes the tool_id used by the toggle tools. add_tool_to_toolkit is idempotent — safe to call if already attached.
  • Custom SQL tools can only be toggled here, not authored. Creating or editing a custom tool's SQL, parameters, or name is admin-UI only — list_toolkit_tools returns admin_ui_create_custom_tool_url and list_custom_tools returns admin_ui_edit_url; relay those.
  • Never pick inputs for the user. Which toolkit, which connection, which ops to expose are the user's choices — when unspecified, list the options (via the list_* tools) and ask.

Wiring a toolkit into Claude Code

create_toolkit / list_toolkits return mcp_remote_server_url. To use the toolkit's tools, register that URL as an MCP server the same way as Setup above (PAT + Basic, or OAuth), then query it through connect-ai-base. (The connect-ai-manage skill's mcp-command helper can assemble the claude mcp add line, base64-encoding username:PAT, if the user prefers.)

What this skill does NOT do

Route these to connect-ai-manage (direct REST) — the Management MCP has no tools for them:

  • Workspaces and data assets
  • Jobs — cache jobs and scheduled queries
  • Users, roles, invites
  • Personal Access Tokens (mint / revoke)
  • Billing — subscription and usage
  • Hard deletes of connections/toolkits, and custom-tool SQL edits (admin UI / REST)

For querying or writing data in any connected source, route to connect-ai-base (+ the connector skill), or connect-ai-direct when no MCP connector is available.