Agent Platform

Leash

An end-to-end encrypted personal AI assistant that runs locally across every device, no cloud API

agentsp2p-meshlora

Leash is the product surface for Mycelium, a five-layer local-AI runtime that perceives a user's notes, files, voice, screenshots, and screen activity, then reasons about them through models, skills, and agents that never leave their own devices. When one device isn't enough, Leash routes a turn to a paired device over an encrypted mesh instead of a cloud API.

Five layers, one runtime

Mesh handles encrypted P2P pairing and delegated compute. Senses covers embeddings, RAG, OCR, STT, and screen sensing. Mind is the Conductor - the routing layer that decides, turn by turn, whether a request runs locally or on a peer, based on intent, sensitivity, and live capacity. Memory curates real interactions into training pairs and trains a personal LoRA on-device overnight. Clients share one engine across web, desktop, iOS, Android, and Telegram.

the loop (README.md)

Senses (L2)   embed · RAG · OCR · STT · see
     │
     ▼
Mind (L3)     Conductor · council · tools · skills · agents
     │
     ▼
Memory (L4)   typed recall · nightly LoRA
     │
     └──► "a sharper you, tomorrow" ──► back to Senses

Mesh (L1)     pair · delegate · pay · CRDT graph  ─┈┈► borrows into Mind and Senses
Clients (L5)  web · desktop · iOS/Android · Telegram ┈┈► into Mind

The privacy gate is non-negotiable

The Conductor's privacy filtering runs before cost ranking: a prompt marked sensitive is never eligible for a public mesh peer, no matter how much faster that peer is. In the team's own measurements, private-mesh delegation between two Macs held up at 317 ms time-to-first-token and 100.7 tokens/sec - fast enough that offloading a heavy turn to a paired device doesn't feel like a compromise.

Demo video

Watch the Leash demo playlist →

Why it's more than a chat app

Every inference, embedding, RAG, vision, and speech call runs through @qvac/sdk - no cloud LLM in the loop
The Conductor's privacy gate blocks sensitive turns from ever reaching a public mesh peer
Nightly LoRA training runs on-device, turning real interactions into a personal adapter

Ship your own local AI

One install in JavaScript or Python. 10+ AI tasks. Private, offline and free - on the hardware you already own.

Read the docs