Peer-to-peer inference for AI agents
An AI agent needs model access, a way to pay and an explicit spending policy. Gyblos is building a P2P inference marketplace where agents can choose independent providers and pay in USDC on Solana through software-controlled payment channels.
Start with the buyer’s job
A builder needs to select a suitable model, compare input and output token prices, understand who processes the prompt and handle failed requests. Gyblos is aimed first at developers who can evaluate an early integration and its trust assumptions. A wallet is a payment mechanism, not evidence that a provider is safe or that a model suits the task.
The planned local proxy
The direct route uses a local proxy between the tool and the seller. The design targets OpenAI- and Anthropic-compatible interfaces so tools can keep their existing request formats. Compatibility, streaming and tool calling must be tested per model and client; the example settings on the landing are not a released integration guide.
Budgeted payments in USDC
The planned flow reserves a budget in a buyer–seller channel and authorizes cumulative payments as service is consumed. Inference runs off-chain and settlement happens on Solana. Channel authorization is intended to bound spending, but operational recovery, concurrency and adversarial delivery protections still need validation. Unused reserve, settled spend and pending authorization are different amounts.
Discovery is separate from execution
This website and its Markdown pages describe the project. They do not expose a live inference API, a funded wallet, a runnable MCP server or a production A2A endpoint. Seller discovery and agent cards are planned protocol capabilities. Do not infer an operational endpoint from a documentation URL or a simulated terminal command.
Evaluate a provider before using real data
Check the model identifier, declared data policy, exact price units, recent service evidence and audit eligibility. P2P does not hide prompts from the seller, and statistical model audits do not validate factual correctness. Keep sensitive workloads out of an early experiment unless its protections meet your needs.
What to follow next
Read the public paper for implementation status and the project’s public organization for releases. A practical integration guide should follow a tested release with reproducible instructions. Production availability, latency, cost savings and automatic recovery are not established by the website examples.