Turn repeatable workflows on almost any website into tools your AI assistant can use.
Agent Process is a browser extension that lets you create your own WebMCP tools on top of existing websites, even if the website does not support MCP or WebMCP itself.
Record a process once, define its inputs, and let your AI assistant (ChatGPT or Claude browser extension) reuse it whenever needed.
Demo: agentprocess.org/demo.
Video: YouTube.
AI assistants (ChatGPT or Claude browser extension) are getting increasingly capable at using websites, but they often need to inspect the interface and figure out the same workflow again and again.
Agent Process takes a different approach.
If you already know how a task should be done, you can record that process once and expose it as a reusable WebMCP tool.
Instead of asking the AI to rediscover every click, it can simply choose the right process, provide the required parameters, and let Agent Process execute it.
This can make repetitive browser tasks:
- Faster
- More predictable
- Easier to debug
- More transparent
- Less token-intensive
- Easier to share across a team
And because the process belongs to the user, you are not limited to the MCP functions an application developer decided to expose.
You can create an Agent Process in two ways:
- Record a workflow using the Inline Manual Builder
- Create or import a process definition as JSON
A process contains the individual browser interactions required to complete the task.
For example:
Create a lead
Add a customer
Create a ToDo
Publish an article
Update a CRM record
Processes can accept parameters that are provided when they are executed.
For example, a Create lead process might accept:
name
email
company
The same process can then be reused with different data every time.
Import the process into the Agent Process browser extension.
The extension registers the process as a WebMCP tool directly in the browser.
When an AI assistant calls the tool, Agent Process executes the predefined workflow on the website.
The AI does not need to rediscover the whole interface or determine every individual step again.
One of the main ideas behind Agent Process is that the website itself does not need to change.
You do not need:
- Native MCP support
- Native WebMCP support
- An API
- Access to the website's source code
- Cooperation from the website developer
Agent Process runs on top of the existing web interface.
This also means you can create workflows tailored specifically to how you work rather than being limited to generic tools exposed by the application.
Agent Process uses the Inline Manual Builder as a no-code recorder and editor.
Instead of manually writing the process definition, you can simply perform the workflow in your browser.
The Builder captures:
- Individual steps
- Elements being interacted with
- User inputs
- Browser interactions
The resulting process can then be exported and used by Agent Process.
The long-term goal is simple:
Perform a task once. Let your AI assistant know how to do it forever.
Agent Process keeps an execution history so you can see what happened when a process was run.
This makes automated workflows easier to understand and debug than completely opaque browser automation.
You can see whether a process:
- Completed successfully
- Failed
- Encountered an unexpected step
Browser agents are great when the path to completing a task is unknown.
But many tasks are repetitive and already well understood.
If you already know exactly how something should be done, asking an AI to inspect the page and reason through the entire interface every time can be unnecessary.
Agent Process combines both approaches:
Unknown task → let the AI figure it out
Known task → reuse a proven process
This makes browser automation more deterministic while still allowing AI to decide when a process should be used and what data should be provided.
Imagine you regularly create leads in a CRM.
Instead of asking your AI assistant:
Open the CRM, find where leads are created, figure out the form,
fill in these details, and save it.
you can expose an agent process.
{
"title": "Google it",
"description": "Go through Google",
"steps": [
{
"title": "Click on ",
"element": {
"selector": "#ti6dpd",
"type": null,
"tagName": "TEXTAREA"
}
},
{
"title": "Click on Search with Google",
"element": {
"selector": ".aajZCb .gNO89b",
"type": "submit",
"tagName": "INPUT"
}
}
]
}Agent Process already knows how create_lead should be performed.
The AI only needs to choose the tool and provide the parameters.
Agent Process uses WebMCP to expose recorded browser processes as tools that AI assistants can discover and call.
Our use of MCP is slightly different from a traditional MCP implementation.
Instead of exposing capabilities from an application we own, Agent Process lets users define their own capabilities on top of other applications.
WebMCP running directly in the browser is what makes this model possible.
Agent Process is an early project built for the WebMCP Hackathon.
The current version supports a limited set of browser interactions, and compatibility can vary between websites.
Modern web applications can contain:
- Dynamic elements
- Overlays
- Delayed rendering
- Complex frontend frameworks
- Interfaces that change frequently
These are challenges we already deal with extensively while building Inline Manual, and we plan to continue improving process reliability.
There are also currently some limitations around automatic WebMCP discovery in AI browser clients. For ChatGPT, custom instructions may be required to make sure the assistant looks for and uses the available WebMCP tools.
MCP can tell an AI assistant what an application can do.
Agent Process can tell it how you want it done.
We see Agent Process as a reusable layer between people, AI assistants, and the applications they use every day.
A kind of procedural memory for AI agents:
Learn a workflow once. Reuse it whenever you need it.
Try Agent Process and see the demo at:
Agent Process was created as our submission to the WebMCP Hackathon.
It started as an experiment:
What if anyone could add their own WebMCP tools on top of any website?
We think there is a lot more to explore.
Created by Inline Manual team.