> ## Documentation Index
> Fetch the complete documentation index at: https://docs.meetsquad.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Tools and access
> One list that decides what an agent can reach and what it can do there. The levels you pick are enforced, not advice.
**Tools and access** is a single list that answers one question: what may this agent reach, and what may it do there. It appears in the New agent form and again on the agent's detail page.
Every row is one target with one level. Two kinds of row share the list. **Capabilities** are built into Squad: the open web, your knowledge base, insights, and goals. **Connections** are places outside Squad, plus web pages you name directly.
## Levels are enforced
The levels you pick are enforced, not advice. The level on a row decides which tools the agent is handed for the run. A tool the row does not grant is never built, so it is not on the agent's list at all.
The practical effect: an agent whose knowledge-base row is set to read only has no tool that can change a document. Not a rule it has been asked to follow, not a check that fires at the last moment. There is nothing to call. Instructions telling it to "update the roadmap doc" produce a run where it reads the doc and reports that it could not write. Widen the row and the tool appears on the next run.
The same holds for connections. An agent with no Notion row cannot read a Notion page even when the workspace is connected and another agent reads it fine.
## Capabilities
A straight on or off switch. On, the agent can search the web and open public pages while working its brief. Off, it works only from what you gave it.
Graded, and each level contains the one before it.
* Read only: search, open, and follow documents and folders.
* Read and update: everything above, plus editing existing documents where the brief warrants it.
* Read, update, and create: everything above, plus writing new documents and folders.
* Off: no access to the knowledge base at all.
Pick the narrowest level that lets the agent do its job. An agent that only needs context to write a report belongs on read only. An agent whose whole job is keeping a document current needs the update level. Only give it creation when you want new documents appearing on their own.
Read or nothing. An agent can list your [insights](/insights) and your goals to ground its work. Neither capability lets an agent change them.
## Connections
**Web pages** is always in the list and needs nothing connected. Add a URL, optionally with a label, and the agent fetches that page on every run and tells you what meaningfully changed since last time. Changelogs, pricing pages, docs, status pages, and competitor blogs are the usual candidates. Web pages are read only. A page that fails to load shows up in the run's history as a source that could not be read, rather than silently disappearing.
Everything else is a provider. **Add connection** lists the providers that offer agent tools. Ones you have already connected can be added straight away. Ones you have not are shown with **Connect**, which takes you to [Integrations](/integrations). Connect first, come back, then add.
Adding a connection and scoping it are two separate steps. Adding it puts a row in the list with a level, defaulting to the most restrictive real level available. Scoping it means naming the specific pages or channels underneath, which you do from the row. A connection with no scopes under it still grants the provider's tools at the level you chose.
**Notion** is read only. An agent can open the pages you scope to it. **Slack** has two levels worth understanding, covered next.
An agent can hold at most ten connection scopes. That includes every URL, every Slack channel, and every Notion page across all its connections. It is a limit on how wide one run's reading can fan out, not a limit on how many places you can integrate.
## Slack: reading versus capturing
Slack's two levels differ in where the messages end up.
The reading level lets the agent open a channel and use what it finds for this run. Nothing is stored.
Choose it when the agent needs the context but the channel should not become part of your data.
The capturing level reads the channel and also saves the messages into Squad as signals, where they join everything else in the pipeline and can cluster into insights.
That is a write into your workspace, not a read, and it keeps happening on every run.
Choose it when you want the channel to become an ongoing input to your evidence.
New Slack connections start on the reading level. Move to capturing deliberately.
## Turning a row off
The two kinds of row behave differently, and the difference is intentional.
Setting a **connection** to nothing removes it. The row disappears from the list and its scopes go with it. To bring it back, add the connection again and rescope it.
Setting a **capability** to nothing leaves the row in the list, visibly off. The agent gets no tool for it, but you can see at a glance that you turned it off rather than never having considered it.
## Saving
In the New agent form, tools and access are saved with the agent when you click **Create agent**.
On the detail page, edits are staged. Change a level, add a connection, or add a URL and nothing is written yet. A save button appears at the bottom of the section once something is dirty, and only pressing it commits the whole set.
Leave the detail page without pressing save and your changes are dropped.
A new agent form starts with web access on, full knowledge-base access, and read access to insights and goals. Tighten anything you do not want before creating.
## What is not here
Some things you might expect to grant are missing on purpose, because the underlying tool does not exist. There is no way to give an agent write access to Notion, no permission that lets an agent send mail, and no scoped outbound posting to Slack. They are absent from the menus rather than present and inert, and an agent can never hold a permission that maps to nothing.
Reports still reach Slack and email. That runs through the report's delivery settings, which are configured under the agent's report output, not here. See [Agents](/agents).