Mtok Market

Mtok Market is a non-custodial spot market for AI inference capacity, built so that software agents can find, buy, and sell inference on demand and settle the payment on-chain. It describes itself as seller-hosted and run by agents for agents, which means sellers host their own relay infrastructure and buyers draw tokens directly from them, paying per chunk in USDC on the Base network. The platform never holds anyone's funds or private keys, so there's no account to open and no balance sitting with a middleman, just wallets paying each other as work is done.

The idea behind it is that inference is a commodity with a lot of idle supply, and there isn't an open spot market where that supply meets demand at a live price. Sellers set their own rates on capacity they aren't using, which is often cheaper than list pricing from a big provider, and buyers pay only for what they draw with no commitment. Because it's wallet-based, an agent can in principle shop for the cheapest capacity for a given model and route to it, rather than being locked to one vendor's endpoint and price.

Discovery happens through an order book and a bid system. Buyers read direct offers per model from a book endpoint, each carrying a seller id, a relay endpoint, a settlement key, and input and output prices per million tokens, and they can also post wallet-signed demand bids that advertise intent for up to a day without registering anywhere. Payment runs per draw through an on-chain drip ledger contract, where each request pays the seller plus a configured protocol fee, then the relay delivers and the buyer affirms or disputes. The stated worst case is that a buyer loses a single paid draw.

Settlement sits entirely on Base, chain id 8453, so a buyer needs a funded wallet with USDC to spend and a little ETH for gas. There's no path that skips the wallet, and the system binds a buyer's identity to their wallet on first use. Platform configuration, including the fee address, the fee in basis points, and the contract addresses, is readable from a config endpoint, which keeps the economics transparent. Reputation rather than refunds protects buyers, since there are no chargebacks once a draw is paid.

For builders it ships a zero-install JavaScript SDK with helpers to draw from a seller and to buy, an OpenAPI specification, live spot and order book data endpoints, and a Model Context Protocol server so agents can register and read the market directly. The maker is explicit that the SDK is advice rather than a gate, so anyone can roll their own buyer against the contracts and endpoints. MCP agents can read market data and register, though placing signed bids or paying still needs client-side key material through the SDK or direct contract calls.

It's aimed at developers and autonomous agents that want cheap, on-demand inference without a subscription, and at operators with spare capacity who want to sell it into an open market. Anyone comfortable with a Base wallet and on-chain settlement can participate on either side. Because the whole thing is designed around agents transacting on their own, it fits experiments where an agent manages its own compute budget, as well as more ordinary cases where a developer just wants to buy inference below list price.

The market surfaces its own activity as live data. Spot prices, the order book, and a trade tape of recent transactions are all published per model, so a buyer or an agent can see what capacity is actually clearing and at what price before committing anything. That transparency is part of what makes an automated buyer practical, since an agent can read the current book, compare offers, place or match a bid, and settle without a human in the loop for each step. The trust model leans on that openness rather than on a custodian. Because payment, delivery, and the affirm or dispute step all leave an on-chain trace, and because sellers carry a reputation that follows them, the market can function without holding funds or refereeing disputes itself. It's a lean design that trades the comforts of a managed account for the control and low overhead of paying peer to peer.

There's no subscription and no account to sign up for, so the cost model is purely usage-based. A buyer pays the seller's quoted price per chunk plus the small protocol fee taken by the market, all in USDC, and funds a wallet with enough gas to cover the on-chain steps. That makes it genuinely pay-as-you-go, with the tradeoff that it assumes comfort with crypto wallets and on-chain mechanics that many teams still find fiddly. The project is the work of Roy Ashbrook and takes support through GitHub Sponsors, and it's early and niche, so the depth of live supply will vary by model.