Every Thunk.AI user can create and save connections to other systems. They can be managed from the user’s Account page. Many connection types are supported and more are being added:
Google Drive (cloud-based file system)
Microsoft 365 (cloud-based file system)
SQL databases (all the common flavors of SQL databases)
Any Model-Context-Protocol (MCP) service
Twilio
Github
Any website that requires a user signin
Google AppSheet applications
Once created, these connections can be used within tools used in any thunk owned by the user.
Most of the connection types automatically create tool libraries in the user's account with a pre-built set of tools based on the connection. For example, if the user has a Google Drive connection, they automatically have tool libraries to access Google Docs, Google Sheets, and Google Drive.
Sometimes, the user may want to create custom API tools to enhance the capabilities of the AI agents in a thunk. Instead of hard-coding API credentials inside each tool, the user can reference credentials from the connection.
Cloud File System Connection
The cloud file system connection is very important to be able to create and use practical thunks. When building meaningful thunks, it is very common to read files (content folders), read and write spreadsheets (lookup and write relevant information), and create new files (write out summaries or reports). Therefore, we strongly recommend that every user include a Google Drive or Microsoft 365 file connection with read and write permissions. To add these permissions, go to your Account -> Connections page, choose the specific connection, and click on the appropriate button to add the necessary permissions.
When you create a new account with Google Drive, the system (and therefore your AI agents) automatically receives a very limited access to your cloud file system. This limited access allows the system to read and write _only_ the files that you explicitly add via the Google Drive file picker UI. This allows for basic testing, but usually is insufficient for fully automated operation.
MCP Connection
The Model-Context-Protocol (MCP) is a standard protocol that all AI platforms use to connect AI agents with external applications. For example, if you have a Salesforce instance and you want an AI agent to be able to read and write data from it:
You would use a Salesforce-specific MCP server, which exposes a standard API over the Salesforce data and capabilities that you wish to use. The MCP server has a url endpoint that can be used to access it
You register the MCP server endpoint with you AI agent platform. In Thunk.AI, you do this by adding an MCP connection to your account. The AI agent platform can then act as an MCP client, retrieving the list of available tools and making them available to AI agents to use.
Creating an MCP server from scratch involves writing code and deploying it. In practice, thousands of pre-existing MCP servers have already been created to integrate with most popular systems, and many application vendors are also starting to natively support the MCP protocol. This makes it easy to integrate with a broad range of applications
Website Connection
As an integral part of the standard tool libraries in Thunk.AI, there is an open_url tool used to read web pages. When an AI agent uses this tool, it uses a hosted web browser to load the page and read its content. While this works with any public web page, it does not work by default for web pages that require authentication.
In order to make this class of web pages accessible to your AI agents, you need to create a website connection. When creating such a connection, you will actually open the website in a browser tab and go through the authentication process. After successful sign in, your authentication credentials are persisted in a browser cookie. This is saved and reused by the AI agent to be able to open web pages within the site.
This feature can be brittle because every authenticated website has different rules with respect to cookie timeout. After the authentication cookie times out, you will need to repeat the authentication process. There is no secure way for the AI agent to automatically sign in on your behalf.