Documentation

Watchers

A watcher is the core concept in Vigil. Think of it as a dedicated AI agent assigned to monitor a specific category of email — each running independently with its own memory, instructions, and alert preferences.

You might have one watcher for work, another for finances, another for package deliveries
Watchers don't share knowledge and don't interfere with each other
Each is completely self-contained, billed at actual AI cost + 5%

Creating a watcher

Creating a watcher takes about a minute. You give it a name, write a prompt describing what it should pay attention to, and choose a reactivity level. Once created, Vigil assigns it a unique forwarding address — something like work-a7f3k9@vigil.run.

The prompt is the most important configuration. Some examples of what you might write:

A job applications watcher might be told to track application status and flag interview requests
A finances watcher might be told to watch for unexpected charges and flag missed receipts

The more specific your prompt, the more useful the agent becomes over time.

What happens when an email arrives

The moment an email arrives at your watcher address, the agent wakes up. It doesn't read the email in isolation — it first loads everything it already knows.

Accumulated memories from past emails
Current state of all active threads
Any rules you've set

It then groups the new email into an existing conversation or starts a new thread, analyzes the content, and decides what to do. Most of the time it tracks quietly. Occasionally something crosses the alert threshold and it sends you a notification. Every decision — including the ones where it chose not to alert — is recorded with the agent's full reasoning.

Configuration

Each watcher has a handful of settings worth understanding:

System prompt — your primary lever; describe what matters, what to ignore, and how to handle ambiguous situations
Reactivity — controls how aggressively the agent alerts you, from very conservative (only active security threats) to permissive (anything notable)
Memory sensitivity — controls how much the agent retains between emails; higher means a richer picture of people, patterns, and situations
Silence threshold — how long a thread can go quiet before flagging the inactivity; useful when you're waiting on someone who has stopped responding
Tick interval — how often the agent proactively reviews your threads for stalled situations, even when no new email has arrived
Model — different models have different cost and capability tradeoffs; you can switch at any time

Multiple watchers

You can run as many watchers as you need, and this separation lets you give very different instructions to different contexts.

Your work watcher can be aggressive about deadlines while your personal watcher sits at low reactivity
A conversation in one watcher has no influence on decisions made in another
Each watcher is billed independently at actual AI cost + 5%