Every tool runs itself.
You approve.
A sovereign browser autopilot. WebLLM in your browser watches a tool's state, proposes the next useful action from a whitelist, waits for your approval, then executes. Nothing leaves your device.
How it works
<script type="module">
import { installAutopilot } from 'https://sjgant80-hub.github.io/fall-autopilot-kit/src/autopilot.js';
installAutopilot({
toolName: 'falllegal-us',
systemPrompt: 'You run the desk for a US attorney.',
model: 'Llama-3.2-3B-Instruct-q4f32_1-MLC',
state: () => ({ matters: window._app.matters, calendar: window._app.calendar }),
actions: [
{ id: 'draft_intake', name: 'Draft intake form',
when: 'new matter, no intake',
params: ['matterId'],
run: async ({ matterId }) => window._app.draftIntake(matterId) }
]
});
</script>
A floating "Autopilot" button appears. First click downloads the ~1.9GB model to IndexedDB (one time, works offline after). Autopilot reads state, proposes the next action, you approve, it runs.
Tier ladder
| Tier | What it does | Autonomy |
|---|---|---|
| T1 Autopilot · assisted | Reads state · proposes one action at a time · user approves each | user gates every action |
| T2 Scheduled | Runs on timer/calendar · proposes batches | user gates batches |
| T3 Signed | Ed25519 identity · signed audit trail | user still gates |
| T4 Trusted whitelist | Low-stakes pre-approved · exceptions only | partial |
| T5 Bonded | Two autopilots cooperate · shared workflow | pair-level |
| T6 Swarm | Team of autopilots delegating across roles | swarm-level |
| T7 Fully autonomous | PWA background task · initiates without prompting | full |
Sovereign by design
WebLLM local
Model runs in your browser via WebGPU. No OpenAI, no Anthropic, no server. Cached once, offline forever.
User-gated
Every action shown as a proposal card with reasoning. Approve to run. Reject to feedback the model. Nothing runs without your click.
Whitelist-scoped
The model can only propose actions the tool explicitly exposes. No arbitrary code execution. The manifest is the sandbox.
Auditable
Every proposal, approval, execution, and rejection logged to IndexedDB. Exportable as JSON. Full trace of who approved what and when.
Wave 1 · flagship manifests
Twelve fall* tools ship with hand-tuned action manifests: falllegal-us, fallbooks-us, fallhr-us, falladviser-us, fallclinic-us, fallinsurance-us, fallestate-us, fallmortgage-us, fallrecruit-us, fallclaim-us, fallvet-us, and FallOS as the cross-tool orchestrator. See manifests/ for each.
Install as MCP server
claude mcp add fall-autopilot -- npx -y @ai-native-solutions/fall-autopilot-kit fall-autopilot-mcp