Custom MCP tools
MCP, the Model Context Protocol, lets you wrap any API or service as a tool an Ask agent can call. Use it to extend Ask with anything your business runs on.
Prerequisites
- An Ask account with permission to register or connect tools. This is permission-gated in some orgs.
- An MCP server you can reach, either one you built or one provided by a vendor.
- Credentials and scopes for the tool, if applicable.
One dialog, two entry points
Register a new tool (top-right of /connectors) and Connect a new tool (in the footer card: "Can't find the right tool? Manually connect external MCP or custom tools…") open the same dialog: Register a new tool (MCP server), "Connect an external MCP server to use its tools."
Register an MCP server
-
Open Tools in the sidebar and click Register a new tool (or the footer card's Connect a new tool).
-
Fill in the dialog:
- URL (required): the MCP endpoint, e.g.
https://example.com/api/mcp. - Name: the tool name shown on its card.
- Authentication type: Automatic (default) or Bearer Token ("If the MCP server requires an API key, enable bearer token authentication").
- Category: where its card files in the Tools page.
- URL (required): the MCP endpoint, e.g.
-
Click Setup connection. Ask connects to the server, discovers its tools, and the card appears in your Added section with an auto-generated description. Its individual tools become attachable from the Add tools picker and manageable under Personal settings → Tool permissions.
-
Test the connection by attaching it to a draft agent and trying it from the Test panel.
Credentials & permissions
Custom MCP servers authenticate at the org level with what you enter in the dialog (Automatic or a bearer token). Vendor connectors (Gmail, Slack, …) work differently: adding the tool doesn't share anyone's credentials, and each user is asked to connect their own account when they first add the tool to an agent without being connected. Either way, be conservative about what a server can reach: a narrower surface means a smaller blast radius if something goes wrong.
Troubleshooting
| Symptom | Likely cause | Fix |
|---|---|---|
| Tool doesn't appear in the Add tools picker | Not connected to your account, or not shared | Re-check the Added section in Tools; ask the owner to share |
| Agent never calls the tool | Instructions don't mention it; tool description is unclear | Mention it by name in instructions; tighten the tool's description metadata |
| Tool calls fail with auth errors | Token expired or scopes too narrow | Reconnect; check the requested scopes against the MCP server logs |
| Tool calls fail with 404 or timeout | Endpoint unreachable from Ask's runtime | Confirm the MCP server is publicly reachable (or on the allowed network for on-prem instances) |
Best practices
- Stable schemas. Don't rename parameters once agents depend on them. Version your MCP endpoints if you need breaking changes.
- Clear descriptions. The tool's natural-language description is what the LLM reads to decide when to call. Be explicit.
- Idempotent writes. Where possible, design endpoints so repeated calls are safe. Approvals mitigate accidents but don't eliminate them.
- Observability. Log every tool call on your side. When an agent misbehaves you'll want the trace.
Tips & gotchas
Start with a single read-only endpoint, attach it to one agent, and validate the call shape end-to-end before adding write actions.
A registered MCP tool with broad write scopes, shared org-wide, on an agent with no approval requirement is a recipe for incidents. Layer scopes, approvals, and per-agent visibility.