Intel / article

Bell Personal Assistant Stack

Source: browserbase.com
Author: Browserbase
Date: Published Oct 28, 2026

Context

BELL is Browserbase's reference architecture and sample app for a personal assistant that carries out a request instead of only replying to it. Its announcement follows one text message, "order me a bag of coffee," and names the service that handles each step. Linq carries the conversation, keeping it in one thread even though iMessage and RCS are separate systems with different APIs. Eve is where the developer defines the assistant's instructions and exposes tools, and it checkpoints the task so the work can pause for credentials or payment approval and resume with its intact. Browserbase runs the browser that searches the merchant's site and fills the cart while staying within the user's $30 limit. Vercel has separately added a Linq connector in which incoming messages continue the same eve session, the same pairing of messaging and planning layers that BELL uses.

Most of the post concerns the two steps where a merchant needs to trust the agent. Identity has two parts in its account. Web Bot Auth can help access providers such as Cloudflare and reCAPTCHA identify the company operating an agent, but not the customer it acts for, so the post adds Baselayer, which attaches tokenized identity-verification (KYC) data to requests. Signing in to the customer's account uses 1Password's agentic autofill, which the post says gives the agent scoped access to vault credentials without exposing the secrets to the model. At checkout, Link by Stripe lets the agent send the user a spend request for a merchant and amount; once the user approves it, the agent receives a payment credential scoped to that purchase rather than the card details. Visa's Intelligent Commerce and Trusted Agent Protocol work addresses recognition of agent purchases at the payment network.

The post is a map of which partner service handles each layer, written in Browserbase's terms. Its consistent design choice is that both sensitive steps, signing in and paying, give the agent scoped access instead of the underlying password or card. The post does not show how the sample app wires these services together or how complete it is. That code is in the GitHub repository the post links to, which this explanation does not draw on.

Key takeaways · AI-distilled

BELL assigns each layer to a separate service: Linq keeps iMessage and RCS replies in one thread, Eve holds instructions and tools, Browserbase drives merchant sites, 1Password supplies credentials, Baselayer verifies customers, and Stripe Link pays.
Eve checkpoints execution, so a task can pause to fetch a user's credentials or wait for payment approval, then resume with its context intact.
Per the post, Web Bot Auth helps providers like Cloudflare and reCAPTCHA identify which company operates an agent, not which customer it acts for. Baselayer fills that gap by attaching tokenized KYC data to help merchants fight chargeback and refund fraud.
At checkout, the agent sends a Stripe Link spend request for a specific merchant and amount. After the user approves, the agent gets a payment credential scoped to that purchase, never the underlying card details.

Terms in this piece · Glossary

context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.

Read the source browserbase.com