Add your own MCP server as a connector

The built-in connectors cover the apps most teams live in. When the tool you need is not one of them but publishes an MCP server, register that server yourself and your Talent gains its tools on the next turn. This is the opposite direction from connecting an MCP client: here TalentOS is the client.

Add a server

Open Connections in the console and scroll to Custom MCP servers.
Select Add MCP server, then choose MCP server in the dialog.
Paste the server's HTTPS URL, for example https://tools.example.com/mcp.
Give it a label, or leave it blank and TalentOS names it after the host.
Select Connect. The provider opens its own sign-in and consent screen in a new tab. Approve there, and you land back on Connections.

The row then shows Connected, the scope, and how many tools the server offers. Select Check at any time to re-test it.

Where you add the server sets who can use it. Added from your own Connections page it is Personal, reachable only by your Talent. Added from the admin Connections view it is Workspace, reachable by every member's Talent, and that takes an admin. You can remove your own personal servers; only an admin removes a workspace one.

How your Talent uses it

Your Talent asks the server what tools it has before it uses any of them, so it always works from the live list rather than a remembered one. Then, for every single tool call:

Your Talent parks the call and asks you to approve it.
You see the exact server, tool, and arguments it wants to send.
Nothing runs until you approve, and the approval is bound to the arguments you actually saw.

Custom MCP tool calls are always ask, and there is no setting that makes them automatic. That is deliberate: approving "any tool on any server I ever add" is not a decision anyone can make in advance. If you want a tool to run on a schedule without you watching, use a built-in connector instead. See Connections.

Each approved call runs at most once. If it fails partway, your Talent reports the failure rather than retrying, because a remote tool may not be safe to repeat. Approve a fresh call to try again.

Limits worth knowing

10 workspace servers, and 10 personal servers per person.
Up to 200 tools are read from any one server.
The URL must be a public HTTPS address. Private and local addresses are refused, including after a redirect.
Removing a server deletes it along with the stored credentials. There is no archive to restore from, so re-add it if you need it back.

A worked example

An operations team uses an internal order system that exposes an MCP server. An admin adds its public HTTPS endpoint as a workspace server and completes the consent screen. A member then asks their Talent to look up order 4812. The Talent shows the exact find_order call and order number, waits for approval, then returns the result. The Activity receipt records the approved call.

What to check

Failures show as a short reason on the server's row. What each one means:

"That URL is not a public HTTPS address." You pasted an http:// URL, a local address, or something that redirects to one. Use the provider's public HTTPS endpoint.
"The server did not answer as an MCP server." The URL is reachable but is not an MCP endpoint. Most providers document a path ending in /mcp.
"The server rejected the access token." The sign-in expired or was revoked on the provider's side. Remove the row and add it again to sign in fresh.
"The server could not be reached." The provider is down or slow. Select Check in a few minutes.
No Custom MCP servers section at all. Check that you opened the admin Connections view for a workspace server. If you are already an admin, reload once and contact support.
The section is there but your Talent says it has no such tool. Select Check on the row first. If it reports tools and your Talent still cannot see them, start a new message so it reads the connection list again.