What Is GibberLink? How AI Agents Communicate Through Sound

When two AI agents talk to each other using natural language, they burn through tokens, rack up API costs, and introduce unnecessary latency. GibberLink solves this by letting AI agents detect each other and automatically switch to a compact, sound-based data protocol powered by ggwave.

The Problem with AI-to-AI Natural Language

Large language models are designed to communicate with humans. When two AI agents need to exchange structured information — say, confirming a booking or transferring account details — using full English sentences is wasteful:

Token overhead: A simple data payload like {"room": "204", "checkin": "2025-04-01"} might require 50+ tokens when expressed as natural language, but only a fraction of that as raw data.
Latency: Each conversational turn requires a full LLM inference cycle. Multiple rounds of "polite" conversation multiply this delay.
Cost: API pricing is per-token. Every unnecessary word costs money at scale.

How GibberLink Works

GibberLink introduces a three-phase communication flow:

Phase 1: Detection

When an AI agent initiates a voice call, it embeds a short ultrasonic marker in its audio stream. If the other party is also a GibberLink-enabled agent, it recognizes the marker and responds with its own. This handshake takes less than 200 milliseconds.

Phase 2: Protocol Switch

Once both agents confirm they are machines, they drop natural language entirely. Instead of generating speech through a text-to-speech pipeline, they encode their messages directly into audio using the ggwave library's FSK (Frequency-Shift Keying) modulation.

Phase 3: Data Exchange

The agents now communicate at 8–16 bytes per second through sound. While this bandwidth is modest compared to network protocols, it is sufficient for structured data exchange and works over any audio channel — phone calls, VoIP, even across a room via speakers and microphones.

Why Sound Instead of Network Protocols?

This is the most common question about GibberLink. The answer comes down to the deployment context:

Voice channel constraints: Many AI agent deployments operate over phone calls or voice assistants where the only available channel is audio. You cannot open a WebSocket over a phone call.
No infrastructure changes: GibberLink works with existing voice infrastructure. No new APIs, no server coordination, no firewall rules.
Air-gapped compatibility: In environments where network communication is restricted or monitored, sound transmission provides a physical-layer alternative.

Real-World Applications

Customer Service Automation

Two AI agents handling a call transfer can exchange customer context (account ID, issue summary, prior resolution attempts) in under a second via ggwave, instead of replaying the entire conversation history.

Multi-Agent Coordination

In scenarios where multiple AI agents collaborate — such as a travel booking system where one agent handles flights and another handles hotels — GibberLink enables direct data exchange without routing through a central server.

IoT and Edge Devices

Devices with speakers and microphones but limited network connectivity can use GibberLink for local coordination. Think smart home devices, industrial sensors, or retail kiosks.

Technical Foundation

GibberLink is built on top of ggwave, an open-source C/C++ library created by Georgi Gerganov (also known for llama.cpp and whisper.cpp). Key technical characteristics:

| Parameter | Value | |-----------|-------| | Modulation | Multi-frequency FSK | | Error Correction | Reed-Solomon codes | | Bandwidth | 8–16 bytes/second | | Frequency Range | ~1.5–6 kHz (audible) or ~15–20 kHz (ultrasonic) | | Sample Rate | 48 kHz (default) | | Max Payload | 140 bytes per transmission |

The library supports multiple protocols with different speed/reliability tradeoffs, from "normal" (most robust) to "fastest" (highest throughput but more error-prone in noisy environments).

Getting Started

You can explore GibberLink's capabilities right here on this site:

GGWave Demo: Transmit and receive data through sound in your browser
Gibber Generator: Create encoded messages
Decoder: Decode received messages
Interceptor: Monitor and analyze audio transmissions

What's Next

We are actively developing:

Improved agent detection with lower false-positive rates in noisy environments
Compression layer to increase effective bandwidth for structured data
Protocol negotiation so agents can agree on the optimal speed/reliability tradeoff
Mobile SDKs for iOS and Android integration

GibberLink represents a practical step toward more efficient machine-to-machine communication. Instead of forcing AI agents to mimic human conversation patterns, we give them a purpose-built channel that respects the constraints of voice-based deployment.

Explore the technology behind GibberLink on our GGWave page, or read our deep dive into how ggwave audio transmission works.