If AI agents are going to save consumers time and do things on our behalf, they need to know who we are.
They need to know how we behave IRL
- what choices we make
- where we like to go
- what things we like
and then use that context to make helpful recommendations or decisions automatically.
At Crosshatch, we've shipped the internet's first personalized inference gateway - a way for users to authenticate their personal AI to apps and agents. Today we share a beta version of our API Playground to see our API in action.
The personalization promise gap
Consumer internet companies still talk about what AI will do for consumers, but they’re light on the details on how.
For instance, last week Booking Holdings’ CEO Glenn Fogel sat down with Fortune magazine
But we’re old enough to remember a human travel agent. And that human travel agent knew a lot about you. She kind of knew what you could afford. She kind of knew what you liked in general, and she suggests things for you. That’s what we want, that personalization, that’s what we want to do. And we can do it. We will do it. And the more you work with our services, the more we know about you.
Consumer internet companies have been telling this personalization story
We'll collect a bunch of data and finally be able to deliver services that anticipate your needs!
for at least a decade, but it’s yet to yield the personalization promised.
Why is this so hard? First party data doesn’t work. Businesses don’t observe enough consumer behavior data to achieve Fogel’s aim. We saw this first hand at Walmart – a consumer staples business where people are shopping all the time. First-party data just isn't enough context.
Travel is a consumer discretionary business. People don’t book trips that often. Booking Holdings’ doesn’t have enough touchpoints with a consumer to actually know a user’s evolving preferences.
A user books a trip with Booking, or maybe Expedia, or maybe directly, they travel and then their life changes. All Booking sees is the travel booked with Booking.
Without a change, it will never yield the personalization Booking’s Fogel imagines.
Why we built Crosshatch
This is why we created Crosshatch: To enable users to log in with their personal AI to any application, where this personal AI has complete context of their preferences, behaviors, and needs, sourced from every application a users uses.
Applications like Booking.com offer travel booking and planning services users trust, but it lacks the context to provide the human travel agent experience Fogel imagines. Instead, we believe users should be able to log in with a personal AI and have Booking “call their agent” for advice on what the user might like, etc.,
Our actions across applications produce data that ‘reveal our preferences.’ Rather than having to repeat ourselves to every app, we should be able to bring our own context-rich AI to help save us time.
What makes Crosshatch different?
Unlike other personalization solutions that rely solely on first-party data or invasive tracking, Crosshatch:
- Unifies context across apps - Brings together data from Gmail, Calendar, Plaid, fitness trackers and more for a live stream of context that’s always up to date
- Puts users in control - Fine-grained permissions let users decide exactly what context to share
- Maintains privacy - Applications never directly access raw user data
- Requires minimal integration - Developers or AI agents can add personalization with just a few lines of code.
- Separates model and context layers - Applications can activate context with the best model for the task and not get locked in to any model provider
- Not mere MCP - Crosshatch creates long-running context syncs, session-based auth, a unified context layer, and folds AI into servers to make interop of consumer data and AI privacy-compliant
Introducing Crosshatch Playground
To make this new pattern more intuitive, we've built Crosshatch Playground – a web application to see how personalized inference works.
Our personalized inference gateway is a proxy over secured top language models, but with a few important differences.
Compute to data
Most applications operate under a “data to compute” paradigm, where users must send their data to remote third-party-governed compute. This is how Booking.com works – you must tell them information about yourself, and they use that on their compute to save you time.
This leads to users having to repeat ourselves across applications. It’s also less privacy preserving, in that we have to tell apps practically everything so that they might be able to save us time.
Crosshatch operates under a “compute to data” paradigm, where applications call your agent – private AI with permissioned access to context – enabling applications to perform compute on your context. This way, you provide your data once and apps are able to ask your agent questions about your context.
AI authenticated by user-specific tokens
Unlike most AI endpoints that are authenticated by a developer- or application-specific API key, our AI API is authenticated by app and user-specific authentication tokens, granted by our Link.
This is because our API is a personalized inference gateway. When you talk to Claude or GPT-4o through our API, you’re talking to that model as well as the user’s context.
Context in the header
When a user logs in to an app with Crosshatch, the user links to the application specific context like from their Gmail or Calendar. When this happens for the first time, Crosshatch initiates a live stream of this context to a database we manage on behalf of the user. This makes sure that context is always up to date without any extra work from the user or the app.
Applications access this context through a header in our API, enabling apps to reference recent reservations, purchases, exercise history, likes and more all by an SQL-like querying language. With these context queries defined, applications can inject the results of these queries into AI prompts.
This design makes it easy for apps and agents to use our API.
When Crosshatch receives an API request, we
- perform the specified context query on behalf of the application
- hydrate the prompt with the results
- Send the hydrated request to the specified AI model
- Return the results to the application
This enables applications to get private personalized inference in a serverless fashion.
We created our playground to make it easy to see how context queries are defined and how to inject them into prompts.
AI gateway
Our API is a proxy over top language models. Through our API you can use models from OpenAI, Anthropic, and Gemini using their native syntax.
The playground uses the Crosshatch provider for Vercel’s AI SDK, which makes calling our API easier in web applications.
Try Crosshatch Playground
Ready to experience the future of personalization? Our playground is now available in beta for developers and product teams to experiment with.
Or if you're ready to integrate Crosshatch into your application:
Join us in building an internet that truly understands its users – where your context can be easily and safely linked across applications, saving you time and delivering experiences tailored specifically to you.
