THE ARCHITECTURAL FORK
Today, we face a fundamental choice in how AI is built for humanity. Personal AI computing shouldn't belong to centralized walled gardens. It belongs to individuals who own their hardware, run their own agents, and control their data.
Centralized Cloud Silos
- • Personal data resides on corporate servers
- • High latency, high cost per token
- • Fixed, restrictive API boundaries
- • Single point of failure & privacy loss
- • Vendor monetization of user context
Sovereign Edge + Infinity Burst
- • Agent lives locally on YOUR hardware
- • Encrypted, Secure Enclave protected
- • Local-first: 0ms IPC, zero-copy memory
- • Burst to cluster supercomputers (0 → 1 → ∞)
- • Sovereign user ownership of model & memory
The Missing Link: hu_ssh Protocol
An iPhone isn't just a phone—it's an ultra-high-bandwidth edge supercomputer. Apple Silicon's Unified Memory Architecture (UMA) pushes bandwidth from 150 GB/s on mobile up to 800+ GB/s on Ultra configurations.
The missing link is an open, hardware-rooted protocol that lets your agent live locally for 95% of daily tasks and seamlessly burst from $0 → 1 → ∞ onto cluster-scale NVIDIA supercomputers when heavy multi-agent reasoning or code compilation demands it.
So we designed hu_ssh (hushh-secure-shell) — an open protocol specified to sit alongside Anthropic's Model Context Protocol (MCP) and Agent-to-Agent (A2A) communication. It is published as a specification and a reference design; the consent layer beneath it, PCHP, is real working code in the open-source hushh-labs/hushh-research repository. See what is actually running for the line between the two.
Wire-Level Protocol Specification
hu_ssh utilizes a fixed binary header designed for zero-allocation deserialization in Rust and Swift:
HEADER_FIXED_SIZE constant in both the Rust and Python listings below. The field table sums to 53 (2+1+1+1+8+4+2+2+32), and the Rust parser compounds it by reading PAYLOAD LENGTH from offset 15 where the table puts it at 13. Three numbers, none agreeing. The listings below are left exactly as drafted so the error stays legible; treat the field table as authoritative and the constants as wrong. The specification derives its total by summing the fields, which is how this was found.MAGIC (2B) | VERSION (1B) | FLAGS (1B) | FRAME TYPE SEQUENCE NUMBER (8B) PAYLOAD LENGTH (4B) SIGNATURE LENGTH (2B) | RESERVED (2B) DEVICE ID SHA-256 (32B) PAYLOAD DATA (Variable) SEP SIGNATURE DATA (Variable, Optional)
Frame Types
- 0x01 HandshakeInit — Establish trust and capabilities
- 0x10 AgentCommand — Request execution on remote
- 0x20 ToolCall — Invoke a tool in the sandbox
- 0x21 ToolResponse — Return result to sender
- 0xFE Heartbeat — Keep connection alive
Reference design
Status: proposed, not yet published as code.
The components below are the reference design the specification calls for — the shape each piece takes and the job it does. They are not files you can clone today. The open hushh-research repository contains substantial PCHP consent-protocol work, but it does not yet contain a hu_ssh wire implementation. The path names are the intended layout, written down so the design is legible and so an implementer — us or anyone else — has something concrete to build to. When the code lands, this section will link to it and say so.
Rust Zero-Copy Parser
Frame parsing with BytesMut, zero allocations on the hot path. Intended home: protocol/hu_ssh_core/src/lib.rs.
C-FFI Bridge for iOS
Safe C bindings for Swift integration, with hardware attestation. Intended home: protocol/hu_ssh_core/src/ffi.rs.
Python NemoClaw Daemon
A QUIC server runtime with a skill execution engine behind a kernel-bounded sandbox. Intended home: consent-protocol/.
Hermes Skill Runtime
Persistent execution with skill-library registration, scoped to granted consent. Intended home: consent-protocol/.
Swift Orchestrator with Fallback
Automatic failover from the remote cluster to local MLX on connection loss, so the person never sees an error. Intended home: hushh-webapp/ios/.
Performance Metrics
| Metric | Centralized Cloud | hu_ssh Private | Win |
|---|---|---|---|
| Handshake Latency | ~120 ms | ~18 ms | 6.6x Faster |
| Local Inference | N/A (Internet) | <10 ms | Infinite Offline |
| Sandbox Boot | ~450 ms | ~12 ms | 37x Faster |
| Privacy Guarantee | Cloud Storage | Hardware Enclave | Zero Leakage |
Architectural Invariants
- Bring Your Own Key (BYOK): The user-controlled key boundary stays on the device.
- Zero-Knowledge Persistence: Unencrypted user memory is never stored on servers.
- Scoped Execution: Actions execute strictly within granted consent scopes.
- Tri-Flow Parity: Unified client alignment across Web, iOS, and Android targets.
Join the Open Protocol Movement
Personal AI computing shouldn't belong to centralized walled gardens. It belongs to individuals who own their hardware, run their own agents, and control their data.
GitHub Workspace: https://github.com/hushh-labs/hushh-research
Protocol Endpoint: hu_ssh://spec.hushh.ai
Collaborate: open-source@hushh.ai
Let's make agents truly useful by making them run seamlessly on the supercomputers of today. 🤫