---
title: MCP servers
description: Connect an MCP server once and every agent in every new chat can use its tools. Sign-ins stay on your machine and agents never see your tokens.
sidebar:
  icon: plug-01
---

MCP servers give agents tools for other services: Linear, Supabase, Context7, Notion and many more.
In Glade you connect a server once, in **Settings → MCP servers**, and every agent in every new chat
can use it.

## Connect a server

Pick one from the catalog, grouped by kind (docs, databases, infrastructure, monitoring, design and
more), or choose **Add custom** and paste a server's `https://` address.

Depending on the server, connecting works one of three ways:

| Server needs   | What happens                                                         |
| -------------- | -------------------------------------------------------------------- |
| **No sign-in** | It connects straight away.                                           |
| **OAuth**      | Your browser opens the vendor's sign-in page, then returns to Glade. |
| **A token**    | You paste a token you created on the vendor's site.                  |

## Your tokens stay with Glade

Connectors and their sign-ins are stored in a private file on this machine. Each agent session reaches
a server through Glade, which attaches the credential on its way out. **Agents never see the
token**, and a session can only reach the servers it was granted when it started. See
[privacy](/reference/privacy) for everything else that stays on your machine.

## Choose which agents get it

A server serves every agent by default, including agents you add later. Untick **Use in all agents**
when connecting, or choose **Choose agents** on a connected server, to limit it. Changes apply to
sessions that start afterwards.

## Servers your agents already have

Below your connectors, Glade lists the MCP servers your agents load from their own configuration,
merged by name. When one of them matches a catalog entry, **Use in all agents** connects it once in
Glade. Glade then offers to remove the agents' own copies so they don't load it twice. It always asks
first, and leaves config files with comments for you to edit.

## Limits

- Custom servers must be public `https://` addresses without credentials in the URL.
- Connectors belong to the machine that runs them. An [SSH host](/machines/ssh-hosts) has its own,
  managed from the machine picker in **Settings → MCP servers** and signed in on the host itself.
- Only the machine's owner can add or remove connectors, because sign-in returns to that machine's
  browser. [Paired devices](/machines/remote-access) can use them but not manage them.
- Some vendors only accept pre-registered apps and are left out of the catalog. Their servers still
  work as custom servers if they accept a token.

## Under the hood

Each connector reaches a new agent session as one MCP server named `<name>_connector`, pointing at
Glade's local gateway rather than at the vendor. The gateway checks that session's grant, attaches the
vendor credential and streams the exchange, so the agent only ever holds a short-lived Glade token.

Sign-ins use OAuth with PKCE, and the tokens live in a file only your user can read under the Glade
home. Disconnecting deletes them locally first, then revokes them with the vendor when it supports it.