Connect AI assistants (MCP)

Let Claude and ChatGPT search your Librario library directly — no more dragging PDFs into a chat window or copying search results by hand.

Librario provides a server for the Model Context Protocol (MCP) — an open standard for connecting AI applications to external systems. Your AI assistant is the client, Librario is the server. Once you’ve connected it, it can on your behalf:

• Search in natural language: “Find publications about EN 1993-1-1 steel column buckling published since 2020.”
• See what is new: “What was added to the library last week?” The assistant lists the most recently added entries, regardless of their publication year.
• Look up publication details: “Tell me more about publication #42.”
• Find and browse collections: “Is there a collection on concrete rehabilitation?” Once it has found the collection, the assistant shows what is in it and can also search within that collection alone.
• Download, read, or process attached files — through a short-lived signed link, or directly in the conversation on AI tools that support reading MCP resources. Reading in the conversation works for files up to 1 MB. For larger files, the assistant gets a download link and the publication’s page in Librario, where you can still download the file after the link has expired.

The assistant can also read Librario’s public documentation. Ask, for example, how loan requests work, and it fetches the matching help page, in English by default or in German on request, with the get_documentation tool. The connection needs no extra permission for this. The guide for AI agents describes how assistants keep documentation and holdings apart and how they cite sources.

Quickstart

You need two things:

1. Your personal MCP URL from Librario. Click your avatar in the top right and open Integrations. The page shows whether your plan includes MCP and provides your tenant-specific MCP URL — copy it from the MCP card. You’ll paste this URL into your AI tool in a moment.
2. An AI-assistant account on a plan that supports custom connectors (per-vendor requirements below).

Claude.ai (web — recommended)

1. In Claude.ai, open Connectors settings (under Customize on personal plans, Settings on Team / Enterprise plans) — directly at https://claude.ai/customize/connectors.

2. Click + Add custom connector and paste your MCP URL. Leave authentication on OAuth.

3. Claude.ai opens a new tab to the Librario sign-in page. Sign in. The consent screen then shows the requested permissions.

4. Review the permissions, then click Authorize.

5. Try a search query:

   Find publications about steel column buckling since 2020.

Requirement: a Claude.ai Pro, Team, or Enterprise plan — custom connectors are not available on the free plan.

Claude Desktop (macOS & Windows)

1. Install Claude Desktop and sign in.
2. Open Settings → Connectors → Add custom connector (or in a chat window click + → Connectors).
3. Paste your MCP URL. Leave authentication on OAuth. Click Add.
4. Claude Desktop opens your browser to the Librario sign-in. Sign in, review the permissions, and click Authorize.
5. Try a search query.

ChatGPT

1. Click your profile picture → Settings → Connectors → Advanced and toggle Developer mode on.
2. Back at Connectors, click + Add custom connector.
3. Paste your Librario MCP URL, leave authentication on OAuth, and confirm I trust this application.
4. ChatGPT opens a new tab to the Librario sign-in. Sign in, review the permissions, and click Authorize.
5. Try a search query.

Requirement: a paid ChatGPT plan (Plus, Pro, Business, Team, Enterprise, or Edu) — the free tier does not support custom connectors. On Business, Team, Enterprise, and Edu workspaces the workspace administrator must first enable Developer mode and grant you access before the steps above will work.

The consent screen shows one permission per line, each with a checkbox. For AI assistants there are two: “Search and view publications” and “Download and read attached files”. Untick the files permission and search and detail lookup stay allowed while download links and reading file content stop. Both permissions are ticked by default. Your choice is remembered per connection; a permission the connection lacks returns a permission error to the assistant on the next call.

Connect Microsoft 365 Copilot

With Claude and ChatGPT, each person sets up their own connection. Microsoft 365 Copilot works differently: someone with administrator rights registers Librario centrally once, and the connector is then available to everyone who is licensed for it. Every person still signs in to Librario with their own Librario user the first time they use it, so everyone who searches Librario from Copilot needs a Librario user.

A user comes into being through an invitation in the user management or, with single sign-on with Microsoft, automatically on the first sign-in from Copilot, as long as your plan has free seats. Once all seats are taken, the person sees a notice instead of the consent screen, and your administrators and librarians receive an email. Plan the seats before the rollout.

Microsoft’s Copilot landscape is tangled: the products have been renamed several times, and there are two unrelated ways to connect an MCP server like Librario:

• Federated connector: an administrator registers the Librario MCP URL once in the Microsoft 365 admin center. Librario then appears for every licensed person in Microsoft 365 Copilot Chat, the Researcher agent, Copilot in Excel and Cowork. Each person needs a Microsoft 365 Copilot license. Access is read-only (search, detail lookup, file download); write actions are not possible.
• Copilot Studio agent: if your organisation builds its own agents, you can add Librario as an MCP tool inside a Copilot Studio agent instead. This is more flexible but needs a separate Copilot Studio license. It is aimed at people who develop agents.

With Microsoft 365 Copilot licenses, the federated connector is the obvious path. It is part of advanced AI controls. Copilot Studio is a separate product with its own license. Librario is then available only inside your agent, not in Copilot Chat or the Researcher agent. With Dynamic discovery, Copilot Studio does not need advanced AI controls. Both paths are described below.

Set up the federated connector (admin)

The federated connector does not register itself with Librario. Microsoft asks for a client ID and client secret, so the connector needs an OAuth client issued by a Librario administrator, which is part of advanced AI controls.

1. You need the Global Administrator or AI Administrator role in the Microsoft 365 admin center, plus access to the Teams Developer Portal.
2. In Librario, an administrator creates an OAuth client for the “AI assistants (MCP)” surface, with the callback URL https://teams.microsoft.com/api/platform/v1.0/oAuthRedirect.
3. In the Teams Developer Portal, open Tools → OAuth client registration and add a new registration: the client ID and client secret, Librario’s authorization URL as the authorization endpoint, the token URL as both the token endpoint and the refresh endpoint, and library:read assets:read as the scopes. Leave PKCE enabled. Note the OAuth client registration ID.
4. In the Microsoft 365 admin center, open Copilot → Connectors, choose Create a new connector → Connect to MCP server, and enter a display name, your Librario MCP URL, and the registration ID.
5. Release the connector, optionally via a staged rollout to selected groups first, then to everyone.

Choose OAuth 2.0 as the authentication method. Microsoft’s second option, Microsoft Entra SSO, works only for MCP servers that accept Microsoft Entra tokens. Librario’s MCP server accepts only access tokens issued by Librario itself.

Microsoft changes the menus of this still-young area frequently. For the current, click-by-click instructions, see Microsoft’s documentation: Set up custom federated connectors.

Use the federated connector (end user)

Once an administrator has released the connector, Librario appears as a source in the Researcher agent and in Microsoft 365 Copilot Chat, where you connect it under Settings → Sources. On first use, Copilot signs you in to Librario. The same OAuth sign-in and consent screen as for Claude and ChatGPT. You sign in with your own Librario user, and Copilot searches with your permissions. After that you search in natural language, for example: “Find publications about steel column buckling since 2020.”

Copilot Studio: Librario as an MCP tool

Organisations that build their own agents add Librario directly to a Copilot Studio agent. You need a Copilot Studio license for this.

Copilot Studio offers two kinds of sign-in: Dynamic discovery and Manual. Use Dynamic discovery. Copilot Studio then registers itself. Nobody has to obtain credentials beforehand.

1. In the agent, open Tools → Add a tool → New tool → Model Context Protocol.
2. Enter a server name, a description, and your Librario MCP URL.
3. For authentication choose OAuth 2.0, then Dynamic discovery. Copilot Studio reads Librario’s discovery documents and registers itself. It asks for nothing further.
4. Click Create, then Next. Create a new connection. Librario shows the sign-in and the consent screen. Sign in and click Allow access.
5. Open the tool’s settings under Details → Additional details. Leave the default End user credentials there. Every person then signs in with their own Librario user. With Maker-provided credentials, everyone searches with the user and permissions of the person who built the agent.

For your IT: the person building the agent makes this choice, and Microsoft allows both options by default. To make sure no agent runs on someone else’s Librario user, allow only End-user credentials under Control maker credential options for the environment in the Power Platform admin center. This also covers Power Automate flows an agent uses as a tool. Details: Control maker-provided credentials for authentication. Outside agents, the same applies to Power Automate flows shared with their owner’s embedded connection: they run on that person’s Librario user. The federated connector has no shared sign-in.

Details: Connect your agent to an existing MCP server.

Manual OAuth configuration

Use this path if Copilot Studio asks for a Client ID and Client Secret when you save. Choose the Manual type under OAuth 2.0. Copilot Studio then asks for Client ID, Client Secret, Authorization URL, Token URL and Refresh URL. This path is part of advanced AI controls. Your plan must include them. Dynamic discovery does not need them.

1. An administrator creates an OAuth client for the “AI assistants (MCP)” surface.
2. Switch Require PKCE off while doing so: the manual form in Copilot Studio has no PKCE setting.
3. Copy the Client ID, Client Secret, Authorization URL and Token URL. Enter the Authorization URL unchanged, as the page shows it, without parameters. The page opens only once, and the Client Secret is the only value you cannot get back afterwards. Enter the token URL in both the token URL field and the refresh URL field.
4. Enter library:read assets:read in the scopes field.
5. After clicking Create, Copilot Studio shows a callback URL. If it does not match the one entered at creation, create the client again with that URL.

Connect Google Gemini

With Google the answer depends on the Gemini product, and the most important limitation goes first: Gemini in Gmail, Docs and the other Google Workspace apps cannot connect a third-party MCP server. With a work account, Gemini offers only a curated catalogue of Google and partner integrations, with no setting to enter the Librario MCP URL. Don’t go looking for one there.

These paths do work:

• Gemini Enterprise: a separate, separately-licensed Google Cloud product (not the same as “Gemini for Workspace”). Here an administrator registers Librario as a “custom MCP server data store”. The feature is generally available; in the Business edition of Gemini Enterprise it is still in Preview.
• Gemini CLI: Google’s open command-line tool. It lets an individual connect Librario without any admin involvement.
• The Gemini app with a personal Google account: at gemini.google.com, personal accounts can add a custom app by its MCP URL under Settings → Connected Apps. Google currently offers this only in the US, in English and to people aged 18 or over, and not for work or school accounts. We have not yet tested this path with Librario.

Set up Gemini Enterprise (admin)

Gemini Enterprise does not register itself; it needs an OAuth client issued by an administrator, which is part of advanced AI controls. Setup happens in the Google Cloud console, not the Google Workspace Admin console.

1. Lift the organisation policy that blocks custom MCP server data stores by default, and grant the Discovery Engine Editor IAM role. If your organisation restricts outbound domains or data sources, also allow the Librario MCP, authorization and token URLs.
2. Register Gemini Enterprise as an OAuth client with the callback URL https://vertexaisearch.cloud.google.com/oauth-redirect, and note the client ID and client secret. Both come from the same page as Copilot Studio: OAuth clients for integrations.
3. Under Data Stores → Create data store → Custom MCP Server, create an entry with your Librario MCP URL, the OAuth endpoints (authorization and token URL), client ID, client secret, and scopes. Tick PKCE support.
4. Complete the OAuth sign-in, then explicitly enable the tools you want. None are enabled by default.

For the current, click-by-click instructions, see Google’s documentation: Set up your custom MCP server data store.

Once the data store is set up, Librario is available to users as a source in Gemini Enterprise. They search in natural language, and the agent calls the enabled Librario tools when needed. According to Google, Gemini Enterprise acts on behalf of a user; the scopes define what it may do in Librario.

Use Gemini CLI (developer)

If you work with the Gemini CLI, add the Librario MCP URL as an MCP server in the CLI configuration. The Librario tools are then available to the model in every CLI session. This path needs no admin setup.

Manage and revoke connections

The Integrations page lists every application that accesses Librario on your behalf under “Connected applications”, sorted by last use. Each row carries a type badge, “AI assistant” or “REST API”, plus the name, connect time, last-used timestamp and the granted permissions as individual badges. A connection that has been used, but not in the last 30 days, also carries the badge “Unused for over 30 days”. A connection that has never been used shows “Never used” in place of the last-used time.

Click Revoke on a row to end that connection immediately. The assistant’s next request to Librario returns HTTP 401. The only way to regain access is a new authorisation.