Lobstah: A P2P Large Model Inference Network for Asynchronous Workloads

Lobstah is a federated peer-to-peer (P2P) computing exchange network designed specifically for large language model (LLM) inference. It allows users of consumer-grade hardware like Mac mini to earn credits by contributing idle computing power, and consume those credits to call the computing power of other nodes in the network when needed. Key features of the project include: using the Nostr protocol for node discovery and building a decentralized ledger via signed receipts; not pursuing low-latency real-time dialogue, focusing instead on asynchronous batch processing workloads, positioning itself as a "cost-effectiveness first" computing solution.

The emergence of Lobstah fills a niche in the LLM inference market. Traditional cloud service providers (such as OpenAI, Anthropic) focus on low-latency single-user dialogue scenarios and rely on high-end hardware like H100. Although consumer-grade Mac mini's inference speed is 3-10 times slower, it has value in program-driven, no human waiting required asynchronous workloads. Typical applicable scenarios include: night research agents (literature review generation), batch document processing (PDF summarization), multi-agent collaboration systems (internal calls for CrewAI/AutoGen), synthetic data generation, code review robots, game NPC simulation, etc.

Lobstah uses a modular architecture, with core components including:

• Protocol Layer: @lobstah/protocol (identity and signed receipts), @lobstah/ledger (receipt logs and balance calculation);
• Computing Engine: @lobstah/engine-ollama (Ollama adapter), @lobstah/worker (OpenAI-compatible HTTP server), @lobstah/router (multi-node routing and failover);
• Discovery and Transmission: Node discovery based on the Nostr protocol (Worker posts announcements to Nostr relays, Consumer pulls node lists), supporting centralized trackers as an alternative;
• CLI Tool: @lobstah/cli for key generation, node management, etc. Node discovery process: When a Worker starts, it can post a signed announcement via Nostr; the Consumer obtains available nodes from the Nostr network; the Router intelligently routes based on model requirements.

Lobstah's economic model implements decentralized settlement based on signed receipts. After a Worker completes computing, it generates an Ed25519-signed receipt that records the number of tokens consumed, accumulated in local logs without the need for a centralized account. Usage example: A research agent needs to process 50 papers to generate a review. The traditional OpenAI method costs $5-20, takes 30 minutes but data leaves the local environment; the Lobstah method distributes the task to the network, with Macs of friends in Vancouver/Berlin and the user's own machine processing asynchronously, completing in 6 hours, and the user repays the computing credits contributed by others via their idle computing power.

Lobstah provides an OpenClaw plugin (in the openclaw-extension/ directory), allowing OpenClaw users to directly use it as an inference backend. Currently, the project is in the pre-alpha stage, but end-to-end testing has been implemented (cross-region streaming, signed receipts, replay protection, multi-node failover, Nostr node discovery). Users can install the CLI via npm: npm install -g @lobstah/cli.

Lobstah represents an alternative AI infrastructure idea. It acknowledges that consumer-grade hardware cannot compete in terms of latency, but precisely targets the niche market of asynchronous batch processing workloads. Through P2P architecture, Nostr discovery mechanism, and signed receipt system, it effectively utilizes idle computing power and provides users with a more cost-effective LLM inference solution. For developers who own a Mac mini, this is a viable way to convert idle computing power into value and is worth paying attention to.