Thread
Chat, assistant, agent
Leo did not start as an agent. It started as a chat box with a good model behind it. What turned it into something else was not a bigger model — it was memory, tools, and a rule about asking permission first.
How it happened
Every release that moved Leo from answering to acting.
-
June 2026
Every model, one subscription
Claude, GPT, Gemini, DeepSeek, Qwen, Grok and Llama behind a single gateway — with a different model assigned to each job: answers, code, vision, transcription.
-
July 2026
The assistant gets a voice
Talk to it instead of typing, streamed through the backend rather than whatever speech engine the browser happens to ship.
-
July 2026
From chat to agent
Rebuilt on LangGraph: one memory, tools it can actually run, an audit trail behind every action, and server-side confirmation before anything touches your data.
-
July 2026
Skills you can attach
Reusable instruction sets — ready-made, or pulled from your own GitHub repository — attached to the prompts and roles you already use.
-
August 2026
The assistant is called Leo
A name, a memory that survives between conversations, tools it can run in other services, and a permission gate on every one of them. Leo knows who you are and what time zone you are in.
-
August 2026
Leo joins the live call
A rule for when interrupting is worth it, and memory of the conversation as it happens rather than a replay afterwards.
-
September 2026
Leo reaches you with the app closed
Push notifications, so a reminder that matters does not sit waiting for you to open a tab.
Leo comes with every plan, including the free one.