// field manual
Falcon Documentation
A clean, minimal guide to setting up and customizing Falcon Protocols. Search and filter topics dynamically.
What Falcon is, and what it isn't.
Falcon Protocol is a multi-agent operating layer for your day. It listens, sees, remembers and acts — but quietly. No chatbot pane begging for attention. No 12 notifications. Just the work, and a soft pulse when something is ready.
Think of it as a small studio of agents that share one memory and one voice. You speak to one. They figure out the rest.
Local-first memory. Your day stays on your machine unless you ask otherwise.
Voice-native. Press one key, say one sentence. Falcon does the routing.
Intentionally slow. No popups. No streaming spam. One result, well-framed.
Get falcon onto your desk.
Falcon is currently invite-only. Once your invite arrives:
curl -fsSL https://falcon.sh | sh
falcon login --invite CODE_123_XYZ falcon start
The first run walks you through one permission at a time — microphone, screen, calendar. You can decline every one and Falcon will still work as a quiet text companion.
Eight agents. One conversation.
Falcon ships with eight specialized agents. You never have to call them by name — the router decides. But knowing what's in the room helps.
Brain — long-running reasoning, planning, multi-step research.
Voice — speech, transcription, real-time conversation.
Eyes — screen understanding, OCR, visual context.
Hands — automation: emails, calendars, file moves, scripts.
Memory — your second brain. Notes, decisions, recurring people.
Curator — tabs, reading list, inbox triage.
Coder — diffs, refactors, terminal flow.
Quiet — focus mode. Mutes the others on a schedule.
Talk to it like a person who is paying attention.
Hold ⌥ Option and speak. Release to send. Falcon transcribes locally, routes the intent and replies in the same medium you used.
⌥ + hold → speak ⌥ + tap → ask quietly (text only) ⌥ + ⇧ → private mode (no logs)
How falcon remembers without becoming creepy.
Memory is opt-in per surface. Anything you mark as remembered lands in a local vector store and a plain-text journal you can read with any editor. Nothing leaves your device unless you connect a sync provider.
Edit, redact, or wipe any entry from the Memory pane.
Set a TTL per topic — 'forget this in 7 days'.
Export everything as Markdown, any time.
Let falcon do the boring parts.
Automations are written in plain English and stored as small, inspectable recipes. You can read them, edit them, share them with a friend. No drag-and-drop swamp.
recipe: morning
when: weekday 08:30
do:
- summarize unread in inbox (max 5 lines)
- draft replies for anything tagged 'fast'
- open today's calendar in a sidebar
- play focus playlist at 30%What leaves your machine, and when.
Local-first. The router runs on-device.
You choose which tasks are allowed to use cloud models.
Per-app permissions, revocable at any time.
No telemetry on by default. Ever.
Hook your own tools into the protocol.
Every agent exposes a small, typed RPC surface. You can write a recipe in TypeScript, Python or shell. Below: a minimal example.
import { falcon } from '@falcon/sdk'
await falcon.brain.summarize({
source: 'inbox:today',
max_lines: 5,
})
await falcon.hands.draft({
to: 'mom@home',
tone: 'warm',
about: 'thursday dinner',
})that's the manual.
The rest, you'll learn by living with it. Ready to unlock full access?