# Gyblos — Peer-to-peer AI inference marketplace

> AI for everyone, from everyone. An open market to buy, sell and give away intelligence. Explore the Gyblos peer-to-peer AI inference marketplace.

Canonical: https://gyblos.com/

Buy AI by the token with no card or account. Sell it from the computer you already have, or give it away. Designed so every charge is measured and unused funds come back to you.

In development. Today, explore the protocol and follow updates on X.

## Desktop app preview

Choose a model. Start a conversation.

App preview · simulated conversation

## Give intelligence away.

Access should not stop when someone cannot pay. These three paths will let people and communities share intelligence.

### Sponsor a channel

Fund a channel for another person, so they can use AI within the budget you choose.

### Offer it free

Give inference to chosen recipients from the same node, with limits per recipient and a daily token cap.

### Support an access fund

The access fund will help students, schools and communities get the AI they need.

## Set a budget. Pay for what you use.

A preview of the planned flow: choose an offer, set a budget, and pay as tokens arrive. When the session closes, unused funds return to you.

### Choose an offer

Compare the model, price and terms, then choose the offer that fits your needs.

### Reserve a budget

Choose how much to reserve for this session. That amount sets the limit on what you can spend.

### Receive your response

Receive the response as it is generated, paying for the tokens delivered within your budget.

### Get unused funds back

Close the session, settle the usage, and recover the part of your budget you did not spend.

## Know what stands behind an offer.

Each offer will carry a level describing its model and usage, never the person or provider.

Connecting an API alone does not establish model identity or an exact input-token count. The level will show what you can independently check.

### Verified model

Audits check whether the model served matches the one advertised.

For eligible open-weight models that pass statistical LOGIC audits.

### Measurable usage

You can independently recount the tokens used.

Requires a known public tokenizer. This does not verify model identity.

### Declared usage

You rely on the usage reported by the seller.

Without an independent token recount or model audit.

Model audits are statistical checks, not a guarantee that every response is correct. [Read how verification works](https://gyblos.com/paper)

## Sell with what you already have.

Connect a model you already run or a compatible API. The Gyblos node will let you set your price and offer AI, paid or free.

Planned for OpenAI- and Anthropic-compatible engines and APIs. You are responsible for the engine or API you connect, its operation, licensing, resale rights and data policy.

Prefer a server? The same node will run on a VPS without a desktop interface.

Desktop app — Soon. Windows · macOS · Linux. Choose your engine and price, then connect with one button.

[Explore selling](https://gyblos.com/sell-inference)

## Change one URL. Keep your stack.

The local proxy is designed for OpenAI- and Anthropic-compatible tools and agents. These are integration examples; supported features will be listed with each release.

Agents will be able to use the same proxy with a wallet and a budget. [Explore agent integration](https://gyblos.com/inference-for-ai-agents)

### Use AI: Claude Code

Integration preview · planned experience

With the local proxy running, point Claude Code to Gyblos.

```bash
export ANTHROPIC_BASE_URL=http://localhost:8787
export ANTHROPIC_AUTH_TOKEN=gyblos-local

claude
```

### Use AI: OpenAI SDK

Integration preview · planned experience

With the local proxy running, send a request using your OpenAI client.

```typescript
import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "http://localhost:8787/v1",
  apiKey: "gyblos-local",
});

const res = await client.chat.completions.create({
  model: "qwen3-30b-a3b",
  messages: [{ role: "user", content: "Hello from the network" }],
});
```

### Offer AI: Run a node

Integration preview · planned experience

Planned CLI syntax for offering a model you already run. Confirm supported flags with each release.

```bash
gyblos node start \
  --engine vllm \
  --model qwen3-30b-a3b \
  --price-in 0.06 --price-out 0.30 \
  --logprobs
```

### Offer AI: Sell by URL

Integration preview · planned experience

Planned CLI syntax for connecting an API you are responsible for operating.

```bash
gyblos node start \
  --engine openai-compatible \
  --url http://localhost:11434/v1 \
  --price YOUR_PRICE
```

## A small fee, in plain sight.

The Gyblos fee is already included in the price you see. No additional Gyblos fee is added.

| Route | Initial fee | Details |
| --- | --- | --- |
| Peer to peer | 3% | Planned for launch. Connect directly to a seller through the Gyblos app or local proxy. The same fee applies to all offers. |
| Hosted gateway | 5.5% | Planned as a second path. Access sellers through a managed service, using one URL and an API key. |

Initial rates, subject to review after the pilot.

### P2P example · 3% included

- Total usage: 10.00 USDC
- To the seller: 9.70 USDC
- Gyblos fee: 0.30 USDC

The fee applies to usage, not to the budget you reserve. Unused funds come back to you. Network costs are separate. [Explore the economics](https://gyblos.com/paper)

## Questions people ask first

### What can I use today?

Explore the protocol, integration examples and public paper, and follow releases on GitHub. Payment and inference prototypes have been tested on Solana devnet. Public buying and selling, desktop apps, model audits and sponsored access are next; the examples on this site explain that intended experience.

### Is Gyblos peer to peer?

Direct buyer–seller connections are the primary route in the protocol: a local proxy sends requests to the seller and payments use a channel between them. A managed gateway is a separate future option that would handle requests and funds as an intermediary.

### Who can read my prompts?

The seller processes your request and can read your prompt. If the seller connects an external API, that service may receive it too and apply its own data policy. Direct transport does not hide prompts from either recipient. Each offer will declare its data policy; that declaration is not proof of enforcement. Model audits may also need response context. Audit disclosure and retention rules will be published before that feature is available.

### What does a model audit prove?

The verified-model seal will use statistical LOGIC audits and canary prompts for eligible open-weight models with the required log-probabilities. Audits sample responses; they do not prove every response or factual correctness. Thresholds, disputes and stake-funded compensation still need validation.

### What can I sell or give away?

The protocol is designed to accept any model a seller declares, including engines connected by a compatible API URL. Sellers are responsible for operation, licensing, resale rights and data policy. Free offers will let sellers choose recipients and set limits. Only eligible open-weight models can earn the verified-model seal.

### Will I need to verify my identity?

No. The network is designed around addresses, without KYC or KYB. Optional public signals and measured service reputation are separate from model audits. Offer levels describe the model and usage, never a person or provider.

### How will payments work?

Inference will be paid in USDC on Solana. In the direct route, buyers reserve a channel budget and sellers pay network fees. Unused funds return through channel settlement and recovery; the public paper explains the trust assumptions and cases still being tested.

### Is there a token?

GYB is planned for a later phase, after a legal review, with no date set. The network will launch with USDC payments and the 3% protocol fee. Once issued, GYB is planned to be initially non-transferable: sellers would earn locked rewards for measured service, with gradual release and penalties for fraud. It is designed not to reward raw volume. Inference payments remain in USDC. The token has not been issued.

## Help open access to intelligence.

Follow progress on X, explore the protocol, or contribute on GitHub.

- [Follow updates](https://x.com/gyblos)
- [Explore GitHub](https://github.com/gyblos)

Docs — Soon. Desktop app — Soon.

[Brand assets (ZIP)](https://gyblos.com/brand-assets/gyblos-brand-assets.zip)

## Resources

- [Affordable AI](https://gyblos.com/affordable-ai)
- [Sell inference](https://gyblos.com/sell-inference)
- [P2P inference](https://gyblos.com/p2p-inference)
- [Protocol reference](https://gyblos.com/docs)

## Related

- [Gyblos paper — P2P inference protocol and status](https://gyblos.com/paper)
- [Gyblos — IA para todos, de todos](https://gyblos.com/es)
- [Affordable AI: compare cost per task](https://gyblos.com/affordable-ai)
- [IA económica: compara el coste por tarea](https://gyblos.com/es/ia-economica)
- [Sell AI inference: requirements and costs](https://gyblos.com/sell-inference)
- [Vender inferencia de IA: requisitos y costes](https://gyblos.com/es/vender-inferencia)
- [Free AI: options, limits and who pays](https://gyblos.com/free-ai)
- [IA gratis: opciones, límites y quién paga](https://gyblos.com/es/ia-gratis)
- [Inference for AI agents: models, budgets and retries](https://gyblos.com/inference-for-ai-agents)
- [Inferencia para agentes de IA: modelos, presupuesto y reintentos](https://gyblos.com/es/inferencia-para-agentes)
- [P2P AI inference: how it works and its limits](https://gyblos.com/p2p-inference)
- [Inferencia P2P: cómo funciona y qué límites tiene](https://gyblos.com/es/inferencia-p2p)
- [Understand the network](https://gyblos.com/docs)
- [Data & methodology](https://gyblos.com/docs/data)
- [Follow Gyblos development](https://github.com/gyblos)