tether

A sovereign transport protocol for finite networks.

This is the reference implementation of Tether.

It is a connection-oriented, datagram-agnostic networking stack designed to operate deterministically across all reference frames: from 100GbE datacenter fabrics to high-latency deep-space links. It rejects the "Best Effort" and "Open World" assumptions of the IP stack in favor of Explicit Capacity and Cryptographic Identity.

Architecture

Tether treats bandwidth as a currency and latency as a physical constant.

Explicit Capacity (Grants)

Transmission is transactional. A node MUST NOT transmit data unless it has been explicitly granted the capacity to do so by the receiver (via the GrantAmount field).

• Congestion Control: Congestion is impossible because data is never sent unless buffer space is pre-reserved.
• Flow Dynamics: Eliminates ACK storms and "Slow Start" heuristics by requiring prepaid receipts.

Lamp Mode (Discovery)

To bridge the void between disconnected nodes, Tether defines a mechanism for Speculative Transmission called a Lamp.

• The Initiator sets the LAMP flag (0x01) and expends its own credits to fund the Receiver's ability to reply.
• This enables "Cold Start" discovery without violating the "Finite Universe" constraint.

The 64-Byte Block

The protocol is physically structured as a sequence of 64-byte Blocks (matching the standard cache line size).

• Branchless Header: The first Block (Header) has no opcodes. Every field (Grant, NackSlot, Timestamp) is evaluated in O(1) time for every packet.
• Control Tags: Complex logic (Manifests, Definitions) is segregated into specific "Control Blocks" on Flow 0.

Hardware Acceleration

Tether is designed for mechanical sympathy.

This repository targets these initial architectures:

1. Generic EAL (AF_XDP / Netmap / RIO): The best "OS-Native" fast path available on the host platform (Linux, BSD, Windows respectively).
2. Intel E800 Series (Columbiaville): As above, but natively utilizing DDP (Dynamic Device Personalization) and DPDK (Data Plane Development Kit) instead of AF_XDP, for ideal performance:
   • E810-CQDA2
   • E830-CQDA2
3. NVIDIA ConnectX-7 (BlueField): Utilizing switchdev and hardware flow steering for ARM64 environments.

At this time, there does not seem to be a viable analogue to the above for macOS on Apple Silicon. For now, you should use OrbStack to bridge through a Linux VM that can use AF_XDP. When and if this changes, this repo will implement the relevant approach.

Implementation Status

Experimental

• Protocol Spec: RFC 0001
• Language: Zig 0.16+ (nightly)
• Driver Models: Intel E800-series NICs (Linux) with Hardware Flow Steering.

Roadmap

• Struct Layout: Validating extern struct alignment for the 64-byte Header Block.
• DDP Profile: Generating the .pkg file to teach the E810 parser about EtherType 0x88B5.
• Credit Accounting: Performant implementation of the atomic grant-subtraction logic.
• Control Plane: Implementation of the Flow 0 Tag parser (Hail, Define, Manifest).

---

© 02026 Wilson Bilkovich