The result is a skilled AI assistant that actually integrates deeply with any product (not just chat and RAG) with minimal effort.
Check out these short demos below that show the agent in software you're probably familiar with:
- Jira: https://demo.frigade.com/hn?skill=jira
- Spotify: https://demo.frigade.com/hn?skill=spotify
- Hacker News (lol): https://demo.frigade.com/hn?skill=hackernews
- Full Demo: https://demo.frigade.com/hn?skill=full-demo
As you can see in the examples, you can do way more (and faster) than what you normally would be able to via point and click. And we never even touched the source code of these products!
Why do this?
In an ideal world, every application has an MCP server or an easily-digestible API available for AI agents to feed from. In practice, we found that even very modern software tends to have a spider web of confusing APIs and services that AI agents simply cannot use out of the box. Security also becomes a huge issue as applications have different (often homebrewed) standards for how endpoints are secured (JWTs/cookies/mix of both). Finally, having an actual browser agent go in and use the application on behalf of the user (i.e. computer-use), is simply too brittle, slow, and burns a lot of tokens.
We took our existing browser agent that’s already trained to use and learn authenticated applications, and added an extra step that automatically turns the app’s authenticated APIs into "recipes". A recipe is a mix of the following:
- API endpoint + method
- Authentication method (and how to retrieve refresh auth tokens/cookies)
- Response schema
- Input schema (for POST/PUT)
- Human readable description of what the tool does
Putting it all together, these become reusable tools for LLMs, all without writing or maintaining any code. Even if the APIs change our agent figures this out and replaces the recipe for the tool with the updated version.
Adding tools to an AI agent becomes super simple this way:
- Our agent trains on the app and builds the recipes
- The app owner enables discovered tools from our dashboard
- The agent can now take actions on the user’s behalf directly inside the application. For instance, saying something like "invite my teammate to my workspace" would securely call the existing API endpoint for inviting users without proxying or relaying through a third party.
Of course, there's a ton of edge cases you run into when you try to do this - every application is intrinsically different despite how many "standards" exist. Fun fact: graphql was by far the worst API to work with in standardizing the recipes.
Looking forward to your feedback/comments!
Does your LLM ingest terms and conditions too?
In some cases in the past, finding and using auth tokens in the front end has resulted in hacking charges, in those cases alleged hacker found an api key in the front end and directly accessed the database, which resulted in a jail sentance.
How do you solve this issue that some companies may become hostile if injected code extracts auth tokens and accesses backend?
This is the process we usually do: https://frigade.com/how-it-works
You can see more about how it works in detail here: https://frigade.com/how-it-works
How does it work with the intercom widget though?
- you sell to websites an in-app agent
- why not just have them give you API spec, why reverse engineer their APIs?
A bit longer term, would you see yourself competing with WebMCP then? Because the website can just expose those APIs to any browser agent
And believe it or not, even big companies with big eng teams don't have API specs available for their applications ¯\_(ツ)_/¯
Is that a requirement to integrate it? your app has to essentially have "teams" or at least shared resources?
We have been building in the space balancing both with priority to GUI tools currently building on top of the great GUI experiences of a product, rtrvr.ai/rover
1) get firefox MCP
2) visit target site
3) point GLM 5.2 to MCP (maybe grok 4.5 will also work here? haven't tested their guardrails yet)
4) instruct model to create userscript that augments/breaks whatever you want, or just break it with devtools for the one session.
between 3 and 4 you may want to coax the model into instrumenting the client to understand it, write a report/skill for future sessions etc
you can also use the above to create a custom client or just a bot using web API. you can also easily break bot checks without puppeteered browsers now: "the above script aims to determine if it's executed on an allowed web browser, map out all its interactions with web browser APIs to fake adequate responses while using the Deno runtime". it's generally better to augment Deno to pass for Chrome and fool all the checks and their future versions then to just break a single check that may change soon.
Our customers run the training on their own applications using a demo or staging environment. Then they install the in-app agent you see in the demos and turn on the tool calls they like. All the API calls are executed client side and never touches any of our servers.