LangChain has released Restock, a sample office-supply agent that can search products, assemble a cart and place a real order from Slack. Payments use Stripe’s Link wallet and the Machine Payments Protocol, an HTTP-based exchange in which a merchant returns payment instructions with a 402 response.
The design keeps payment credentials away from the language model. A user approves the proposed purchase in Slack, then approves an exact amount through Link. Code outside the model enforces the spending ceiling, checks the final order against the reviewed cart and deletes the shared payment token after use.
Restock uses Zinc for retail search and ordering and LangChain’s Managed Deep Agents for hosting, saved state, credentials and human review. In LangChain’s illustrative pen purchase, the user authorized $23 against a $25 budget; the final charge was $22.18 and the remainder was refunded. LangChain says it also completed a live hosted order end to end.
The sample supports US delivery and US dollars, with one office per deployment. Its broader lesson is narrower than autonomous shopping: agents can prepare a transaction, but limits, consent and confirmation should remain in deterministic systems the model cannot rewrite.