MindooDB introduces WebMCP to let AI agents use 12 different apps

Moving tools into the browser

MindooDB and Haven have introduced a way for AI agents to interact with 12 different applications, including a CRM, mail client, and task manager. Instead of the agent trying to guess where to click on a screenshot, it calls specific functions. This method allows the model to perform actions like adding a chart or drafting an email as if it were a program calling a function.

The core technology is called WebMCP, which moves the tool logic directly into the web page. Traditional Model Context Protocol (MCP) servers often run as separate local programs or web services that require their own login and API keys. WebMCP avoids this detour by registering tools directly within the browser tab via a specific API.

Because the tools live in the page, they run under the user's current session and the page's own rules. This means there is no need for extra API keys or separate servers for the agent to access data. As mindoodb.com reported, the agent only does what the page already allows the user to do.

The model as a universal translator

Integrating different applications is usually expensive because every pair of apps needs a unique interface and maintenance. With this new system, the AI model acts as the translator between these disconnected tools. It can read a meeting summary in one app and automatically create tasks in another based on the decisions made.

Haven serves as the broker in this process to manage how the agent sees these tools. It provides a set of core tools that let the agent understand where the user is and what apps are currently open. The agent can then request to open a specific app or arrange workspace tiles to complete a task.

To keep things organized, the tools are named based on the specific app they belong to. For example, a tool for a task app might start with 'vega_', while a CRM tool might start with 'cairn_'. This naming scheme helps the model identify which app's capabilities it is currently using during a conversation.

Layered tool discovery

Providing an AI model with too many options at once can cause it to become confused or make mistakes. To solve this, Haven reveals tools to the agent in layers as the task becomes more specific. The agent starts by seeing only the basic Haven tools, which explain how the system behaves and what the user is currently doing.

As the agent begins a task, it can request more information about specific apps. It might call a summary tool to see what an app is for before the system opens it. This prevents the model from being overwhelmed by a long list of available tools all at once, keeping the context window clean and efficient.

The system also prioritizes domain operations over simple clicks. Instead of telling the agent to 'click the checkbox in row 3,' the tool is designed to 'complete_tasks.' This allows the model to work with plain words and high-level instructions rather than struggling with the underlying code or internal storage fields.

Safety and testing methods

Security is a major focus of this implementation, especially since agents are working with real user data. Haven requires users to switch on agent tools in their preferences and give consent for each individual app. Tenant admins can also set signed policies to decide which models are allowed to access specific data.

The developers also implemented several rules to ensure the agent remains helpful and safe. For instance, files are never sent directly to the model; instead, the agent receives a reference to a file in an exchange folder. The system also filters out hidden text from HTML emails so the agent does not see private instructions intended for humans.

To ensure these tools actually work for an AI, the team used a coding agent to test them. They set up test environments where a model had to perform realistic tasks, like turning minutes into tasks. If the model hesitated or made a mistake, the developers used the transcript to improve the tool descriptions until the model could complete the task successfully.