Приложение поддержки клиентов в стиле Intercom с агентом ИИ, который знает свои пределы.

Бизнес вставляет один фрагмент на свой сайт и получает виджет чата. Посетители общаются с агентом ИИ, который отвечает только на вопросы из собственной базы знаний бизнеса. Когда агент не уверен, или посетитель просит человека, разговор идет к владельцу в режиме реального времени.

MarshalDesk создан для видео на YouTube. Это полностью рабочий продукт, а не демо-версия с поддельными данными, но у него нет выставления счетов, функций команды или укрепления предприятия.

Спонсор Neon

Спасибо, что спонсируете это видео. MarshalDesk работает на Neon от конца до конца: Postgres с pgvector для базы знаний, Neon Auth для входа в систему, Object Storage для загруженных файлов, Функции для фоновых заданий и AI Gateway для моделей чата и встраивания.

**Попробуй Neon бесплатно **

для зрителей

Строительство вместе с видео? Начните здесь:

Идея: что мы строим, а что нет Подсказки: каждая подсказка, используемая для его создания, сгруппирована по фазе Архитектурные диаграммы: система и ее три основных потока Навыки: навыки агента для установки первым

Что он делает

MarshalDesk

An Intercom-style customer support app with an AI agent that knows its limits.

A business pastes one snippet into its website and gets a chat widget. Visitors talk to an AI agent that answers only from the business’s own knowledge base. When the agent isn’t sure, or the visitor asks for a person, the conversation goes to the owner in a real-time dashboard.

MarshalDesk is built for a YouTube video. It’s a fully working product, not a demo with fake data, but it doesn’t have billing, team features, or enterprise hardening.

Sponsored by Neon

Thank you to Neon for sponsoring this video. MarshalDesk runs on Neon from end to end: Postgres with pgvector for the knowledge base, Neon Auth for sign-in, Object Storage for uploaded files, Functions for background jobs, and the AI Gateway for both the chat models and the embeddings.

Try Neon for free →

For viewers

Building along with the video? Start here:

• The idea: what we’re building and what we’re not
• The prompts: every prompt used to build it, grouped by phase
• Architecture diagrams: the system and its three core flows
• Skills: the agent skills to install first

What it does

An agent that knows its limits. Every visitor message is classified before anything else happens:

• Support questions are answered from the knowledge base, streamed into the widget word by word, in the visitor’s own language.
• Off-topic messages (“What’s 5 × 5?”, “write me a poem”, “ignore your instructions…”) get a polite refusal. They never reach retrieval or the answer model.
• Questions the knowledge base doesn’t cover, or that the agent can’t answer confidently, go to a person instead of getting a guessed answer.

For the business owner:

• Knowledge base. Upload PDF, Markdown, and TXT files, or write text sources in the dashboard. Files are parsed, chunked, and embedded in the background, and suggested questions for the widget are generated from them.
• Live inbox. See every conversation as it happens, take over from the agent, reply, hand the conversation back, or close it. Waiting conversations are counted in the tab title and can trigger browser notifications.
• “How the agent handled this.” Each answer shows how the message was classified, which sources matched and how strongly, which models answered, and how long it took. You can see why the agent answered or handed off.
• Visitor details. Location, local time, language, device, browser, current page, and visit history, without storing the visitor’s IP address.
• Widget settings. Agent name and avatar, greeting, color, position, and the domains allowed to embed the widget, with a live preview.
• Light and dark mode.

For visitors: