Asking for access to Linear, Gmail or Notion — what the agent does, where you have to click Authorize, and where the credential ends up
“Get into my Linear” is the one ask your agent cannot finish on its own. It can
find out what is connectable, it can tell you exactly which entry to open, and it
can wire the result into itself — but the step in the middle, where somebody
proves they are you, is yours.That step is one screen: Tools → Browse, pick the service, press Connect.
Say which kind of access you mean, because that sentence has two readings and the
agent picks the wrong one. Asked exactly Get into my Linear., the agent on the
account this page was written from read it as a browser task: it opened the
agent-browser skill, started a desktop session and navigated to
https://linear.app/login, saying “Opening Linear on the desktop browser —
that’s where sign-ins belong.” That is a real thing it can do, and it is not
what was wanted.The two asks that are not ambiguous:
Connect the Linear MCP server so you have its tools.Log into linear.app in your browser and read my open issues.
The first is this page. The second is the browser, and it needs your Linear
password, in a live desktop session you watch.
Tell you what is already connected. It asks the platform, because nothing in
its own repository lists them. The answer includes the platform’s own packages
and every server you have connected.
Tell you what a connected server can do. The Telegram Bot server on the
account this page was written from exposes 68 tools; the agent reads their names
and descriptions before it decides which to reach for.
Give itself the tools. The connected server has an id, and a line of the
agent’s own declaration names it. That is a change to the harness like any
other, so it publishes as a version.
What it cannot do is hold your Linear password or click through Linear’s consent
screen. The catalog has 83 entries and 54 of them are OAuth: for those, the
browser has to be yours.
The catalog. Each card carries the two things that decide what connecting
feels like: oauth or manual for the credential, http or sse for how
Splox talks to the server.
2
Pick the service and press Connect
For an OAuth server there is one field, and it is only a label:
Linear: a name for the connection, and a redirect to Linear to authorize
The page says it plainly — You will be redirected to Linear to authorize
access. We never see your password. Gmail (as Google Workspace) and
Notion work the same way.For a manual server there is no redirect. The form asks for exactly the
headers that server authenticates with, named one by one:
Telegram Bot: manual auth, so the form asks for the one header the server wants
3
Tell the agent it is done
It shows up under Your MCP servers with a status of connected.
The server after connecting: on the account, with its credential held by the platform
Now the ask the agent can finish: “I connected Linear — give yourself its
tools.” It looks the server up, reads what it exposes, adds the id to its own
declaration and publishes.
Not in your harness. The platform holds it, attached to your account, and uses it
when a run reaches that server. What the harness carries is the server’s id — a
uuid that means nothing without the account it belongs to.That matters in three ways worth knowing:
The harness is safe to read and to share. A version of it never contains a
token. Someone reading your repository sees "01a06250-80ad-7e53-91b3-…" and
learns only that an agent uses a server you connected.
Disconnecting is one gesture. The bin icon on the card removes the server
and its credential. Anything pointed at it then fails, so an agent naming that
id has to be pointed somewhere else — another ask.
It is per account, not per agent. Connecting once makes the server
available to every agent you have; which agents actually get it is the tools
list in each declaration, and an agent is only ever offered what its own line
names.
This is a different screen from Connections, which is for model provider keys
and environment secrets. Rule of thumb: a service your agent should call goes on
Tools; a key your agent’s own code reads from the environment goes on
Connections → Secrets. See Connections.
Add MCP server takes a name, a URL, a transport and whatever headers the
server authenticates with. It then behaves like any other entry: it appears under
Your MCP servers, it gets an id, and an agent gets it by naming that id. The
details are on Your own MCP servers.If the service has no MCP server at all, the answer is usually not a connection.
The agent has a shell, a real browser and a package manager, so “read my Linear
issues” can also be an ask for a tool of its own that calls Linear’s HTTP API with
a token you put in Connections → Secrets. That is Give it a
tool, and it is the right shape when the thing you want is
narrow.
The card says connected and, on the wire, has_credential: true. A server
with a credential that has expired says so on the same card.
Ask the agent to use it. One call is a better check than any status field —
“ask Linear who I am” costs a few seconds and either works or gives you the
provider’s own error.
The version list shows the publish that gave the agent the tools. If it is
not there, the agent edited its checkout but did not push, and a new chat will
not have it. See Versions.
Two independent things, and you can do either without the other.
What you want
What to do
The agent should stop using it
Ask. “Take the Linear tools off yourself.” One line out of the declaration, one version.
The service should stop being reachable at all
The bin icon on the card. The credential goes with it, for every agent on the account.
Removing the connection while an agent still names its id does not break the
publish — the id is just a string in a list — but calls to it fail at the moment
they are made. Do both, in whichever order, and the loose end is gone.
