Manage MCP servers
See which MCP servers your teams use, and push approved servers to their devices from one place.
The MCP Servers page shows which MCP servers people use and which ones are installed on their devices. Admins can also add managed servers, which the Tokenize desktop app installs on the selected devices.
Who can do this: Admins add, edit, and remove managed servers. People with usage access see the whole organization. Everyone else sees only their own devices and usage.
The tabs
| Tab | What it shows |
|---|---|
| Recently Used | Servers used in the selected date range, with Users, Tool calls, % failed, and Last seen. |
| On Devices | Servers found on devices, with Team, Surfaces, Users, Devices, and Last seen. |
| Managed | Servers your organization pushes to devices. |
If you don't have usage access, the tabs cover only you. Recently Used shows "What you used recently." On Devices lists the servers on your own devices. In page search, this tab is called My Devices. Managed is read-only.
With usage access, you can filter the whole page by team with the picker at the top right.
Read a server's details
Click a server on Recently Used to open its details: Connections, Failures, Tool calls, Tool failures, Users, and Last seen. Some servers are only seen through their tool calls, so Connections and Failures show "-" for them. A note under the numbers says why:
| What you see | What it means |
|---|---|
| "This is hosted on Claude.ai. We see it when tools run; connection counts only apply to servers installed locally." | A Claude.ai connector. |
| A Codex badge on the row, and "This is a ChatGPT connector that Codex calls. We see it when tools run; Codex reports no connections for it." | A ChatGPT connector used from Codex. Its name ends in "(Codex)". |
| A ChatGPT agent badge on the row, and "This is a ChatGPT connector that a ChatGPT workspace agent calls. We see it when tools run; agents report no connections for it." | A ChatGPT connector used by a ChatGPT workspace agent. Its name ends in "(ChatGPT agent)". |
| A note that some or all of the tool calls "came from OpenAI's Compliance Logs" (Codex) or "came from ChatGPT workspace agents" | Those calls are already in the totals. Tokenize reads them through your organization's OpenAI Compliance key. See Connect OpenAI. |
| "The most common error on those failed calls was FORBIDDEN: the service refused access to what Codex asked for, so the person needs access to it there." | Some of a ChatGPT connector's calls from Codex failed. The note names the error code behind the most failures and, when Tokenize knows it, what the code means. Other meanings point to a sign-in to redo in ChatGPT, the service's rate limit, or a mistaken call by the agent. |
| A note that some or all of the tool calls "came from Claude.ai chats and remote Cowork sessions" | Those calls are already in the totals. Tokenize reads them through your organization's Anthropic Compliance API key. See Connect Anthropic. |
| Cloud-agent sign-in errors, and a line such as "12 cloud-agent sessions couldn't sign in" | Cursor cloud agents couldn't sign in to this server. Cursor reports only the sign-ins that fail, so this is a count of sessions, not a failure rate. Sign the server in again where your cloud agents get their MCP settings in Cursor, or remove it if the agents don't need it. This needs Cursor's OpenTelemetry export. |
Add a managed server
- Go to MCP Servers and open the Managed tab.
- Click Add Server.
- Enter a Name. To pick from a list, click the arrow (Browse known servers). The list includes servers already running on your devices and common public servers. Picking one fills in the rest of the form. Names use letters, numbers, hyphens, and underscores.
- Choose a Connection type:
- Hosted server: enter a Server URL (HTTPS only). Add Headers if the server needs them.
- Local command: enter the Local command and any Command arguments (Optional), one per line. Add Environment variables if the server needs them.
- Under Send to, pick one or more teams. Leave it empty to send to every device.
- Click Add Server.
Header and environment variable values go to every targeted device. Don't use personal credentials. Use OAuth when the server supports it.
Managed servers reach the selected devices within 15 minutes. They install in Claude Code, Cowork, Codex, and Cursor wherever those apps are present. People may need to restart their apps before the server shows up.
To change a server, click its edit button in the Managed table, then Save changes. You can't rename a server. To rename one, remove it and add it again.
Remove a managed server
Click the remove button on the server's row, then confirm Remove. This removes the server from every device in your organization on their next sync.
Check delivery
Admins see delivery status at the top of the Managed tab. Each device that has checked in is counted in one bucket:
| Status | What it means | What to do |
|---|---|---|
| Up to date | The device has the latest managed server list. | Nothing. |
| Pending | The device hasn't picked up the latest list yet. | Wait up to 15 minutes. If a device stays pending, check that the desktop app is running and up to date. See Troubleshoot the desktop app. |
| Blocked | The device tried to apply the list but couldn't write to at least one app. | Click Show devices to see which app failed. |
| Stale | The device hasn't reported in a long time. | Check whether the device is still in use. See Manage devices. |
Click Show devices for a per-device view. It shows one column per app (Claude Code, Cowork, Cursor, Codex):
- Applied or Not installed: fine. Not installed means that app isn't on the device.
- Unreadable config: the app's config file couldn't be parsed. Fix or remove the broken file on that device.
- Changed mid-write: the file changed while Tokenize was writing. It usually clears on the next sync.
- Read failed or Write failed: Tokenize couldn't read or write the file. Check file permissions on that device.
The Status column in the Managed table is separate. Active means the server has been seen on at least one device. Pending means no device has reported it yet. Use the delivery status above the table to see how many devices are up to date.
New servers that show up on devices also trigger a New MCP server alert. See Alerts.