Where to find them: Customize
Everything that shapes how the agents work for you is under Customize in the sidebar. It has five sections, one pill each. On every pill but Company, a search box filters the list you are looking at:
Customize opens on Connectors. The old Plugins address,
/plugins, takes you there.
Connecting a connector
Open Customize → Connectors, find the tool under Available, and press Connect. You sign in on the tool’s own site; Attensira never sees your password. The connection is per workspace, and you can disconnect it at any time from the connector’s Manage page. Until a connector is connected, the agents cannot see its tools at all. Connecting is the only way a tool enters their toolset. The list shows what you can connect today, grouped by what each connector is for. A tool we cannot reach, or one we have built but not shipped, is not on the page: a row you cannot use is not an integration, it is a roadmap. If the tool you need is missing, Missing a tool? Request it at the foot of the page reaches a person.Reads and Can
Every connector says, under its name, what connecting it lets an agent read and what it lets an agent do:- Can: nothing yet means the connector is read-only. No agent can change anything in that tool.
- Every write an agent makes waits for a yes. An agent that calls a write stops and asks first, which is why each thing a connector can do says it happens once you approve it. Nothing is sent, posted or published on the strength of the connection alone. The one standing yes is Act: a workspace you switch to Act lets safe hygiene fixes go out without asking each time, and anything that makes a claim still asks.
The page in your workspace is the authority: a connector appears there only when it can actually be connected right now.
Google Drive works inside one folder
After you connect Google Drive, pick the one folder the agents work in. Until you do, every Drive call is refused, reads included, and the row sits under Needs attention with Pick folder. The default for “which of my files may an agent read” is none, not all. With a folder picked:- Reads stay in that folder. An agent can find a file or a folder, read a file’s details, download a file, and export a Google Docs, Sheets or Slides file. Those named operations are the whole list. A search only looks directly inside the folder, and a file named by its id is checked to sit directly in the folder before it is read. Subfolders, sharing, revisions, comments and the rest of your Drive are out of reach.
- Finished work lands in that folder. The one thing an agent can do is save a new file there, after you approve it. It always lands in the folder you picked, whatever folder the agent names. It never overwrites, moves, copies, shares or deletes a file, and never changes a permission.
Google Calendar adds holds, and invites nobody
An agent can read your calendars, their events and your free time, so launches, reviews and calls are planned around time you actually have. The one thing it can do is add an event to your own calendar, once you approve it. Every guest is removed from the request and Google is told to send no updates, so nobody else is emailed or invited. It cannot edit or delete an existing event, or change a calendar’s settings or sharing. Inviting people is outreach, and goes through the outreach path and its limits.Shopify reads products, never orders
An agent reads your products, collections, blogs and shop details. Orders, customers, payouts, discounts and themes are out of reach. What it can change is copy, each change once you approve it:- a product’s or collection’s description, and its SEO title and description. Never its price, variants, title, handle or status;
- a new blog article, always saved unpublished. You publish it from your own Shopify admin.
Needs attention, and reconnecting
Connectors that need you sit at the top of the page under Needs attention, each with the one button that clears it:
When approved work is waiting on a connector, the row says how much: 3 items waiting. Everything else that works is under Connected.
A connection is marked as needing a reconnect when the tool’s provider confirms it no longer works: when an agent’s call through it fails and the provider reports the connection expired, or when the daily check of every connection finds the same. A single failed call on a connection the provider still reports as working marks nothing. When it happens:
- one Inbox row, Reconnect Name, is filed with a Reconnect button. It is not filed again for every failed call;
- work you already approved for that tool is not lost. It stays approved and waits, and it goes out after you reconnect;
- loop health says how many connections need reconnecting.
Where finished work can land
Some connectors are places the agent can deliver to, not only read from. Each one writes under the same approval mode as everything else the agent does outward, and each finished piece shows up in Deliverables with where it lives.
Without one of these, the agent still does the work: it hands you the text to paste, and you mark it shipped once it is live.
Registering your own MCP server
If you run an MCP server, register it under Customize → Connectors and the agent can call it like any built-in tool.1
Add the server
Give it a name and its URL. The name becomes the slug the agent uses to address it. If you already use the server in Claude Desktop or Cursor, paste that
mcpServers configuration instead: each remote server in it is imported with its URL and its headers.2
Give it what it signs in with
An open server needs nothing. Most servers that need a key expect it as a request header, such as
Authorization: Bearer …, which comes across with a pasted configuration. Keys and header values are write-only: the page shows which headers are set, never their values.3
Let discovery run
Attensira calls
tools/list on the server and registers what it finds.command entry in a pasted configuration) cannot be registered: there is nothing for us to reach over the network.
How your tools are named
Discovered tools are namespaced:search tool on a server registered as warehouse becomes mcp__warehouse__search. The namespace makes collisions impossible: two servers can both expose search, and Attensira’s own built-in tools can never be shadowed by a registered one, because the built-ins are assembled first.
Servers that sign in with OAuth
You can register a server whose auth mode is OAuth, but the agent cannot call it yet: Attensira holds no OAuth token for it. The server appears to the agent as one tool that says it needs authorising, so the agent tells you that instead of failing with a 401 nobody can diagnose. UseapiKey or a header instead, or put an authenticating proxy in front of the server.
When a server is unreachable
An unreachable server does not silently vanish from the agent’s toolset. It becomes a single tool namedmcp__<slug>__unavailable that explains the problem when called.
That is deliberate. A tool that disappears looks to the agent like a capability that never existed, and it will confidently tell you it cannot do something it normally can. A tool that answers “this server is unreachable” gets you a real diagnosis.
Limits
Discovery and calls are bounded so one slow server cannot hold up a whole session:
A server that exceeds the discovery budget is skipped for that session rather than delaying the answer.
How connectors relate to Attensira’s own MCP server
These are opposite directions and are easy to confuse.- Connectors and registered MCP servers are inbound to the agents: tools the Attensira agents can call.
- Attensira’s MCP server is outbound to you: it lets your own assistant, such as Claude or ChatGPT, call Attensira. Set it up from Settings → Your assistant; see Connect a client.
What connectors cannot do
- They cannot act without a yes: yours on the item, or the standing one you gave by switching the workspace to Act. Nothing that deletes anything of yours is admitted.
- They cannot be re-read on demand. There is no control that forces a connector to refresh, because there is no endpoint behind it. Rather than show a button that spins and does nothing, the product omits it.
- A connected tool is available to the whole workspace, not to one person or one chat.