You are a model provider now
You can sell tokens on TrustedRouter now, and the seller can be anyone who has an endpoint and a price. You register an HTTPS endpoint that speaks the OpenAI chat API, you set a price per million tokens, and your model appears in the user-provided section with an id that any TrustedRouter customer can call. The model running on your laptop right now qualifies. So does the agent script you wrote last month, the fine-tune sitting on a rented A100, and you personally, answering questions with your hands.
Getting there is one command. Run npx reverse-harness --mode proxy --upstream http://localhost:11434/v1 and the ollama instance you already have starts taking paid requests from everybody holding TrustedRouter credits. Point the same flag at llama.cpp, vLLM, or LM Studio and it works the same way. Swap to npx reverse-harness --mode exec --command "python my_agent.py" and the thing you are selling is your program, running on your machine, with your data on the same disk. The third mode is you: npx reverse-harness --mode human drops incoming prompts into a browser queue and waits for you to type. The client verifies every TrustedRouter signature before it hands you a request, opens a tunnel so we can reach you behind your router, keeps your clock, and answers the health canary itself so nobody has to get out of bed for it. It is Apache 2.0, it is on npm, and it lives at github.com/Lore-Hex/reverse-harness, so you can read all of it before you run any of it.
You set the price the caller pays, and 70% of that price lands in your earnings wallet as TrustedRouter credits, which you can move into any of your workspaces and spend on anything in the catalog. Our 30% comes out of the same number, so the price you publish is the price the buyer sees. Route your own traffic to yourself and you lose 30% on the round trip, which is the correct incentive and stops the obvious game before anyone bothers to play it.
There are three kinds of model you can register, and they differ mostly in patience. A machine model can charge up to $1,000 per million tokens and has thirty seconds to produce a first byte, sixty seconds of idle between bytes, and five minutes on the wall clock. An agent model gets the same ceiling with a minute for the first byte and ten minutes to finish, because agents think and call tools before they talk. Those two are where the volume is: a GPU box in a closet serving a fine-tune, or a retrieval loop sitting next to a corpus nobody is ever going to upload. A human model gets five minutes for the first byte, two minutes of idle, fifteen minutes total, and a ceiling of one dollar per token, which is a million dollars per million tokens. The ceiling is that high because when the model is a person, one sentence can be worth what an hour of that person is worth, and the buyer is a pipeline that would otherwise stop and file a ticket somebody reads on Tuesday.
I went first, with the least serious infrastructure I own, because I wanted to know whether the money would move. My id is trustedrouter/user-joseph-live, the handle on it is jperla, I priced it at ten cents a token, and the endpoint behind it was a cloudflared quick tunnel pointed at the laptop on my desk. Two real requests came in through the production enclave and landed in my terminal. I typed both answers by hand while the gateway held the connection open and streamed my typing back as a completion, and I earned $2.73 in credits, which is 70% of what the caller paid.
Who buys from an endpoint they have never heard of? Nobody, by accident. There is no browsing your way to a user-provided model. You get called by id, so every request that reaches you came from somebody who went looking for you. Two requests and $2.73 is a small pile of money and a large fact: somebody paid a dime a token for sentences a person typed into a text box. The catalog cannot quote that price, because the catalog has no people in it, and it has none of your data in it either. A fine-tune that knows a schema living in four heads, an agent wired into your company's file share, a person who knows why the pipeline broke: those are worth more per token than any commodity model gets paid, and until this week there was no way to charge anybody for one of them.
Serving those calls is bounded work. You POST clock-in, we probe you before we believe you and hand back a 409 if the probe fails, you heartbeat on an interval, and you clock out when you are done. Turning work down is an HTTP error: any 4xx means you looked at the request and judged it, and it costs you nothing at all. Only a 5xx counts against you, because that means you took the job and dropped it, and it takes three consecutive failures of yours to clock you out. A caller who hangs up mid-stream never counts.
The traffic reaching your hardware is fenced on our side. Requests come from an attested enclave over HTTPS with the resolved IP pinned for the life of the connection, redirects refused, and only allowlisted chat fields forwarded, and your credentials are decrypted inside that enclave and nowhere else. Every request carries TR-Signature: t=<unix>,v1=<hex hmac-sha256 of "t." + raw body> with a 300 second skew window, and the harness checks it before your model sees a token. You can tell whether a request is really from us by doing arithmetic.
Buyers get told what they are buying: a user-provided model is not attested and is not covered by zero-data-retention, and we print that on the model page, in the API shape our clients read, and on the public trust page. Getting listed costs you an email on file, one funded top-up of any amount above zero, which is what unlocks phone verification, a phone verified by call or SMS, and then a government ID check through Veriff that wants $25 of lifetime top-ups to attempt and charges $5 against your credits each try. That is enough friction to make a fake seller a losing business and nowhere near enough to stop a motivated one. The morning after my model went live, somebody I have never met signed up, walked every step of that gate alone without asking me a single question, and registered a real hosted Qwen endpoint at forty cents per million tokens, working from nothing but the docs. They were in business before I finished my coffee. Go get paid for the GPU that is idling in your apartment tonight.