AI consultant on the site
launched October 2026Chatwoot + an AI bot that prices products and hands over to a manager when needed. Starting point: ~30% of emails were answered within an hour.
A Telegram assistant for the team: type or dictate a task and the agent does it with its own tools — searches and edits the knowledge base, builds a report on ads and traffic, prepares an ad campaign, fills a product card on the site and uploads photos, answers questions about orders from the OMS.



A person types or dictates a task in Telegram; an LLM agent carries it out with its own tools and replies in text or voice. The knowledge base is a folder of Obsidian notes the owner edits from a laptop and a phone: it also holds user profiles (the file itself grants access), each person's memory and plan, skills and ad-campaign drafts.
Recurring tasks are written into the plan and the bot runs them on schedule and sends the result. A new ad campaign is only prepared by the agent — a human publishes it with a button.
[TBD: translation]
The agent has 51 tools. Shown all at once, every request would carry dozens of schemas and the model would lose its way choosing. Tools are grouped into named bundles; a bundle granted in the profile is visible from the first step, a bundle granted through a skill opens only after the agent itself opens the SKILL.md — the instruction and the tools arrive together. The access boundary is the grant itself: a bundle not granted is never even assembled.
[TBD: translation]

[TBD: translation]

[TBD: translation]
[TBD: translation]