How to give an AI agent access to your knowledge base

The prompt window is not a knowledge base

When you paste a document into a chat window and ask a question, you are doing the retrieval work yourself. You found the file, decided it was relevant, copied the text, and handed it over. The AI did the easy part. This works for one-off questions, but it falls apart the moment you want an agent to do something useful on its own, like summarise yesterday's support tickets, flag conflicting information across your docs, or draft a weekly update from project channels.

An AI agent that runs autonomously, on a schedule or in response to a trigger, cannot wait for someone to paste context into a prompt. It needs persistent access to your team's knowledge base: the documents, conversations, emails, tickets, and files that contain what your organisation knows. Without that access, the agent is limited to its training data and whatever you remembered to include.

What an AI agent needs to be useful

There is a difference between an AI assistant that answers questions when prompted and an AI agent that works independently. An agent needs three things that a chat window does not provide.

Persistent access across sessions. A chat resets when you close it. An agent needs to reach the same sources every time it runs, without someone re-authorising or re-uploading anything. If your product specs live in Notion, your customer conversations in Slack, and your contracts in Google Drive, the agent needs standing access to all three.

Cross-app reach. Knowledge does not live in one place. Teams spread information across dozens of tools, and the boundaries between those tools are where context gets lost. An agent that can only search one app is not much better than that app's built-in search. The value comes from connecting information across sources, finding the Slack thread that contradicts the wiki page, or the email that updates the spec.

Granular permissions. Not every agent should see everything. A customer-facing agent should not have access to internal financial documents. A reporting agent might need read access to your project tracker but should never modify it. The permission model matters as much as the connection itself.

Why API-only approaches fall short

The obvious technical answer to "give an agent access to your data" is to wire up APIs. Connect the Slack API, the Google Drive API, the Notion API, write some glue code, and point your agent at the resulting data.

This works in theory. In practice, it creates several problems.

Each API is its own integration. You need authentication, rate limiting, error handling, and data transformation for every source. A team with fifteen tools needs fifteen integrations, each maintained separately. When an API changes, which happens regularly, someone has to fix the integration before the agent breaks.

The data stays siloed even after you connect it. Pulling results from five APIs gives you five separate sets of results. The agent still has to figure out how to combine a Jira ticket, a Slack conversation, and a Google Doc into a coherent picture. Without a unified index, search across these sources remains shallow.

Most teams do not have the engineering capacity to build and maintain this infrastructure. The people who would benefit most from AI agents, operations teams, support teams, managers, are rarely the people who can write and deploy API integrations.

How Fabric connects agents to your workspace

Fabric takes a different approach. Rather than asking you to build integrations one at a time, it provides a persistent memory and context layer, a vault that sits between your tools and any AI model. You connect your apps once, and every agent you create can draw on that shared knowledge base. Because Fabric is the context layer rather than the model itself, it works with Claude, OpenAI (including ChatGPT), Gemini, DeepSeek, xAI/Grok, and others. You can access your Fabric content from ChatGPT or from Fabric's own AI assistant, choosing the model that suits each task.

The connection process is deliberate. When you add a source, you choose what the agent can access: specific channels, folders, projects, or entire apps. Fabric lets you approve agents, whether Instinct, Muse, or Claude, for limited-time, scoped access to specific tools. An agent might be granted permission to check your Gmail and read one Slack channel for 24 hours, with no access to anything else. You control the scope at the level that makes sense for your team and your use case.

Your content lives in Fabric's cloud drive, which streams files on demand with zero local storage, or remains in the connected services where it already sits. Either way, it is indexed and available to agents through a single retrieval layer.

Once connected, Fabric indexes your content and makes it available through semantic search. This means agents do not just match keywords. They understand what a question is about and find relevant information even when the exact words differ. A question about "customer churn" can surface a document titled "retention analysis" or a Slack thread about "users who cancelled."

With more than fifty connections available, most teams can bring their entire working context into one place without writing any code. The connections page lists the full set, from common tools like Slack, Google Workspace, and Notion to more specialised platforms.

The vault: a secure gateway between your tools and any AI

Most approaches to giving AI agents access to your data involve handing credentials directly to the model or the application hosting it. Fabric works differently. It positions itself as a persistent vault and gateway layer that sits between your connected tools and whatever AI you choose to work with. Your data flows through Fabric, not through the model provider. The model never holds your credentials, never stores your files, and never maintains a direct connection to your apps. Fabric does all of that, and the AI sees only what Fabric decides to surface for a given query.

This architecture means you can switch models, use multiple models in parallel, or let different agents use different providers, all while maintaining a single set of connections and a single security boundary. Your Gmail stays connected to Fabric once. Whether an agent powered by Claude, a query from ChatGPT, or a task handled by Fabric's own AI assistant needs to reference an email thread, the access goes through the same controlled path.

The vault also serves as a persistent memory layer. Chat sessions with AI tools are ephemeral by default: close the window and the context is gone. Fabric retains what matters. The documents it indexes, the memories it builds, the relationships it maps between concepts and sources, all of this persists across sessions, across models, and across time. An agent that runs on Tuesday can build on context that was gathered the previous Friday, without anyone re-uploading or re-explaining anything.

No other tool occupies this position. Individual AI providers offer their own file upload and retrieval features, but those are locked to a single model and a single session. Fabric is model-agnostic and session-agnostic. It is the layer that makes your knowledge base durable and portable, regardless of which AI you point at it.

Scoped, time-limited agent access

When you approve an agent to work with your connected tools, Fabric does not grant open-ended, permanent access. Every approval is scoped to specific tools and limited in duration.

The process is deliberate. When an agent needs to perform a task that involves your connected apps, you choose which tools it may use and for how long. You might approve Instinct (Fabric's autonomous agent) to check your Gmail for messages matching a particular query, giving it access for four hours. You might grant Muse (Fabric's creative agent) read access to a single Slack channel for the duration of a research task. You might let Claude access your Fabric filesystem to scan a set of project documents, with that permission expiring at the end of the day.

Each of these grants is narrow. The agent cannot wander into tools it was not approved for. It cannot retain access beyond the window you specified. And you can revoke access at any point before the time expires.

This model reflects how organisations already think about access control in other contexts. A contractor gets a building pass that works for specific floors during specific hours. A vendor gets access to one system for the duration of a project. Fabric applies the same logic to AI agents: minimal privilege, bounded duration, explicit approval.

The practical effect is that you can let agents do useful, autonomous work without worrying that a misconfigured permission will expose your entire organisation's data. An agent approved to read your product roadmap in Notion cannot also browse your HR documents in Google Drive. The boundaries are set at the point of approval, not buried in a settings page you configured months ago and forgot about.

Accessing Fabric from ChatGPT, Claude, and other AI tools

Fabric is not limited to its own interface. Through its support for the Model Context Protocol (MCP), you can access your Fabric content, your Fabric memories, and your connected tools from external AI applications.

MCP is an open protocol that allows AI tools to query external data sources in a structured way. Fabric implements MCP so that tools like ChatGPT (through its desktop application or plugins) and Claude's desktop app can reach into your Fabric vault and retrieve information. The direction here is worth emphasising: this is not about using ChatGPT through Fabric. It is about using Fabric through ChatGPT. You sit in your preferred AI tool, ask a question, and the answer draws on everything Fabric has indexed across your connected apps.

Setting this up involves connecting your Fabric account as an MCP data source within the external AI tool. Once configured, you can ask ChatGPT a question like "what did the product team decide about the pricing model last week?" and ChatGPT will query Fabric, which in turn searches across your Slack, email, documents, and any other connected source to find the answer. The external tool handles the conversation. Fabric handles the retrieval.

This works because Fabric has already done the hard work of connecting, indexing, and unifying your data. The external AI tool does not need its own integrations with Slack, Gmail, or Notion. It needs a single connection to Fabric, and Fabric provides access to all of them through one retrieval layer.

Claude's desktop app can connect to Fabric through MCP in the same way, giving it access to your full Fabric knowledge base during a conversation. You can ask Claude to summarise a project's status, and Claude will pull from your indexed Slack threads, Google Docs, and Notion pages without you needing to copy and paste anything into the chat.

The MCP integration also means that Fabric's memory engine is available to external tools. Memories that Fabric has built over time, the relationships between people and projects, the evolving context of ongoing work, all of this enriches the responses you get from whatever AI you happen to be using.

Practical examples

A few scenarios illustrate how these features work together in practice.

Checking Gmail for a specific query. You are preparing for a meeting and need to know whether a particular client has responded to a proposal. You approve Instinct to access your Gmail for two hours with a task: find any messages from that client's domain in the past week that mention the proposal. Instinct searches your indexed email, surfaces the relevant thread, and produces a summary. Once the two hours pass, it no longer has access to your email.

Giving Claude temporary access to Slack. Your team has been discussing a product decision across several Slack channels over the past month. Rather than scrolling through hundreds of messages yourself, you grant Claude access to three specific channels for eight hours. Claude reads through the indexed conversations, identifies the key arguments and decision points, and produces a structured summary document in your Fabric workspace. The access expires that evening.

Scanning your Fabric filesystem for a report. You have a collection of quarterly reports, design documents, and research notes stored in Fabric's cloud drive. You approve Muse to scan these files and produce a creative brief that synthesises themes across the past six months of work. Muse reads the files, identifies recurring patterns, and drafts the brief as a document you can review and share.

Querying Fabric from ChatGPT via MCP. You are mid-conversation in ChatGPT, working through a complex analysis. You need to reference a decision that was made in a Notion document and discussed in a Slack thread. Instead of leaving ChatGPT to find and paste that context, you ask the question directly. ChatGPT queries Fabric through MCP, retrieves the relevant Notion page and Slack messages, and incorporates them into the conversation. Your flow is uninterrupted.

Combining scoped access with scheduled tasks. You configure a weekly agent that runs every Monday morning. It has standing approval to read your team's Jira board, your Slack standup channel, and a shared Google Drive folder. Each Monday, it produces a status report summarising what was completed, what is in progress, and what is blocked. The report appears as a document in Fabric, ready for your weekly meeting. The agent's access is limited to those three sources and it cannot reach anything else in your workspace.

Setting up an agent with the right permissions

Creating an agent in Fabric involves defining what it should do, what it can access, and when it should run. The permission model is layered so you can start narrow and expand as you build confidence.

Tool permissions determine what actions the agent can take. An agent might be allowed to search your knowledge base and create documents but not send emails or post to Slack. You decide which capabilities each agent has, and you can change them at any time.

Approval gates add a human checkpoint. For sensitive actions, you can require that the agent's output be reviewed before it takes effect. The agent does the work, produces a draft or a recommendation, and waits for someone to approve it. This is particularly useful when you are still learning what an agent can do reliably.

Scheduling controls when the agent runs. Some agents should run daily, others weekly, others in response to specific events. A reporting agent might run every Monday morning. A monitoring agent might run every few hours. The schedule is part of the agent's configuration, not something you have to manage externally.

The agent produces real documents as output, not just chat messages that disappear. These documents live in your Fabric workspace, where you can review, edit, share, and reference them. This makes the agent's work tangible and auditable. You can see exactly what it produced and decide whether it met your expectations.

Making retrieval work at scale

Giving an agent access to your knowledge base is not just about connecting data sources. It is about making sure the agent can find the right information when it needs it. This is where retrieval-augmented generation comes in.

When an agent runs a task, it does not read your entire knowledge base from start to finish. It formulates a search based on what it needs to accomplish, retrieves the most relevant pieces of information, and uses those to generate its output. The quality of the result depends heavily on the quality of the retrieval.

Fabric's memory engine improves this over time. As you use the system, it builds a richer understanding of how your information connects. Concepts that appear across multiple documents, relationships between projects and people, the evolving state of ongoing work: all of this becomes part of the context the agent can draw on.

This is also why a unified knowledge base matters more than individual app connections. When your agent can search across Slack, Google Drive, Notion, email, and your project tracker simultaneously, it finds connections that no single-app search could surface. The support ticket that references a product decision made in a Slack thread three months ago, the onboarding document that contradicts the current process: these cross-references are where the real value lies.

Reaching agents through the tools you already use

Not every interaction with an agent needs to happen inside a dedicated workspace. Fabric Tag lets you invoke agents directly from Slack or Discord. Tag the agent in a channel, ask a question, and it searches your connected knowledge base to answer.

This matters because it removes a context switch. The person asking the question stays in the tool they were already using. The agent does the retrieval work in the background, searching across all connected sources, and brings the answer back to where the conversation is happening.

Each agent also gets its own email inbox, so you can forward messages to it or include it in email threads. The agent processes what it receives and can respond, file information, or trigger a follow-up task depending on how you have configured it.

For teams that want to integrate agents into their own tools or workflows, Fabric supports the MCP protocol, providing a standardised way to connect AI agents to external systems.

Security and data handling

Giving an AI agent access to your knowledge base raises reasonable questions about data security. Where does your data go? Who can see it? What happens if something goes wrong?

Fabric uses AES-256 encryption and holds CASA Tier 2 certification. Your data is encrypted at rest and in transit. The security page covers the full set of controls, including how data is processed, stored, and deleted.

The permission model described earlier is part of the security picture. By limiting what each agent can access and requiring approval for sensitive actions, you reduce the blast radius if something goes wrong. An agent that can only read your product docs and write to a specific folder cannot accidentally leak financial data or modify customer records.

Frequently asked questions

Do I need to upload all my documents manually?

No. Fabric connects directly to your existing tools. Once you authorise a connection, your content is indexed automatically. You do not need to export, convert, or upload files one at a time.

Can an agent access information across multiple apps at once?

Yes. This is one of the core benefits of using a unified knowledge base. An agent can search across Slack, Google Drive, Notion, email, and fifty-plus other connected apps in a single query.

How do I control what an agent can and cannot see?

When you set up an agent, you define its permissions at the source level. You can grant access to specific channels, folders, or projects rather than giving blanket access to an entire app. You can also restrict whether the agent can read, write, or take actions in each connected source.

What happens if an agent produces something incorrect?

Agents produce documents and outputs that you can review before acting on them. Approval gates let you require human review for sensitive actions. The agent's work is visible and auditable, so you can catch errors before they have any downstream effect.

Can I use Fabric agents without replacing my existing tools?

Yes. Fabric sits alongside your existing tools rather than replacing them. You keep using Slack, Google Drive, Notion, and everything else. Fabric connects to them and gives your agents access to the information they contain.

How is this different from using ChatGPT or Claude with uploaded files?

Uploading files to a chat window gives the AI temporary access to a small set of documents for one session. Fabric gives agents persistent, structured access to your entire knowledge base across all connected apps, with the ability to run autonomously on a schedule.

Can multiple team members share the same agent?

Yes. Agents live in your Fabric workspace and can be shared with your team. Different team members can interact with the same agent through the workspace, through Slack or Discord via Fabric Tag, or by emailing the agent's inbox.

What kinds of tasks can an agent perform with knowledge base access?

Agents can summarise information, monitor for changes, draft documents, answer questions, flag inconsistencies, generate reports, and act within connected apps. The specific tasks depend on how you configure the agent's instructions and permissions.

Is my data used to train AI models?

No. Fabric does not use your data to train models. Your information is used only to serve your queries and your agents' tasks. The security overview provides full details on data handling.

How long does it take to set up?

Connecting your first few apps and creating an agent typically takes minutes. The indexing process runs in the background, so your agent can start working with available content immediately while the rest of your knowledge base is being processed.

Related pages

• How AI agents access your data

• What is retrieval-augmented generation

• How to create a shared knowledge base

• How to build an AI assistant for your team on your own data

• How to use Claude with your company data

• How to make AI search your Slack history

• Knowledge retention solutions

• One search across all your tools