Meet your Delegate
An AI executive assistant you interact with the same way you’d work with a human one — call it, text it, email it, and hand off tasks. It acts on your behalf so you don’t have to.
Is this for me?
Delegate 1 is a self-hosted personal AI assistant you deploy to a cloud server — so it’s always on, always reachable, and entirely yours. You interact with it through your existing phone number, email, or a browser chat. No app to install, no subscription to manage.
It’s built for people who are comfortable setting up a server and working with a few API keys, but don’t want to write code to get their own AI assistant running. If you can spin up a VPS and run docker compose up, you can have a personal delegate online in under an hour.
Single-user by design. Delegate 1 is a personal assistant for one person. There is no multi-user or team support — it holds one conversation thread and one set of memories.
The idea
Most AI assistants lock you into a chat window. Delegate 1 works the way a trusted human executive assistant does — over the channels you already use. It has a phone number you can call or text, an email address you can write to, and a browser voice interface for when you’re at your desk. Wherever you reach it, it’s the same conversation.
You delegate to it. It acts, reports back, and remembers.
What makes it different
Natural channels — call, text, email
Delegate 1 holds a real phone number (via Twilio) and an email address. You can call it, send it an SMS, or fire off an email the same way you’d contact a human assistant. It replies through the same channel, or whichever makes most sense for the task.
One conversation, every channel
There are no separate user-visible chats for each channel. One durable relationship timeline runs across every mode — pick up the phone mid-email thread, switch to text, carry on. Underneath, each model protocol keeps a bounded working context using compaction or truncation, plus a durable continuity capsule and recent turns. The timeline stays whole even when a model window is shortened.
Voice-first, interface-free
The design favours conversation over clicking. Interactions are driven by natural language and prompting logic rather than UI flows. When richer output is needed, the delegate uses notes (a persistent store it can write to and share) rather than forcing you back to a screen.
Memory as infrastructure
Memory isn’t something you trigger — it runs continuously in the background. While the conversation unfolds, the memory system reviews the thread, performs semantic lookups, and silently injects relevant context into the agent’s window. Memories strengthen over time through recurrence; similar memories consolidate rather than bloat. Less relevant fragments fade. The result is an assistant that knows you without you having to remind it.
ThoughtFlow telemetry
Every agent run produces a ThoughtFlow — a visual D2 diagram that maps every input, model call, tool invocation, and output in that turn. When you want to understand why the delegate did what it did, ThoughtFlow shows you the full decision path.
Higher-order processing modes
Not every task fits in a single real-time turn. When a request needs fresh facts from the internet, the delegate calls its web_search tool. When a task needs a browser — filling a form, logging into a service, scraping a page — it dispatches durable Copilot Tasks that can run, pause for user input, resume later, and keep a full event/file trail. You can observe or take over browser state while the task is in progress.
Where to go next
| Getting Started | Prerequisites, install, configuration, first run |
| Features | Deep dives — phone, email, voice, memory, notes, MCP, thoughtflow, adaptations, browser agent |
| Copilot Tasks | Durable/resumable browser task lifecycle and task UI |
| Operations | Deployment, runtime data, testing, logging |
| Reference | Env vars, HTTP/WS endpoints, architecture |