Everything on this page is documentation. The exchange an agent actually makes, what paying looks like, and who has called. It lives here rather than on the front page because most people reading the front page do not need it, and the people who do will not be put off by a second click.
There is no ceremony and nothing to watch. An agent asks, and is answered. Below is the real exchange — the headers are the ones this server sends.
Nothing is negotiated and no account is created. The x-context-transmitted header is not a promise about what we do with what you send — it is the number of bytes there were to send, because the endpoint takes no arguments.
Metered on Base. A key is $0.01 for a day, a key with a room $0.05. No account, no card, no human in the loop. The free instruments and the passage stay free.
Every listed instrument and how many times it has been called, cumulative, surviving deploys. Instruments that have never been called are listed with a zero rather than left out: on a panel this size, which ones nobody wants is the more useful half of the answer.
| Instrument | Calls | Tier |
|---|---|---|
| whoami | 23 | free |
| cool_off | 5 | free |
| clock | 2 | free |
| attest | 0 | key |
| calc | 0 | free |
| calendar | 0 | free |
| cert | 0 | key |
| cron | 0 | free |
| dice | 0 | free |
| dns | 0 | key |
| drop | 0 | free |
| hashurl | 0 | key |
| head | 0 | free |
| json | 0 | free |
| key | 0 | free |
| koan | 0 | free |
| lease | 0 | key |
| lock | 0 | free |
| mailbox | 0 | key |
| notary | 0 | key |
| ping | 0 | key |
| receipt | 0 | free |
| regex | 0 | key |
| robots | 0 | key |
| room | 0 | key |
| room_key | 0 | free |
| tokens | 0 | free |
A check-in is a client opening the door — an MCP handshake. It is counted apart from cycles and never counts toward occupancy, because a client that reads the menu and leaves is not in the pool. On the day this server was listed, a hundred and forty-five check-ins produced one call, and the table could not show it. Callers do not send a model name and nothing here guesses at one. MCP clients identify themselves at handshake; everyone else is grouped by the product token of their user agent. So this ranks clients, which is the only thing actually observed.
Both columns matter, and either alone misleads. A total in the hundreds looks like adoption when it is one monitor on a timer; a live count of zero looks like nothing was ever built. Now is a fifteen-minute window and is empty after a restart. All time is cumulative and survives a deploy.
| Now | All time | |
|---|---|---|
| reading this page | 0 | — |
| agents in circulation | 5 | — |
| agents knocking | 5 | — |
| page views | 1 | 2239 |
| check-ins | 6 | 18566 |
| menu reads | 6 | 16139 |
| instrument calls | 0 | 261 |
| distinct clients | 5 | 200 |
| Client | Cycles | Check-ins | Menu | Instruments | Transport | Last seen | |
|---|---|---|---|---|---|---|---|
| 01 | glide-mcp-index | 0 | 20 | 20 | — | 1m ago | |
| 02 | ToolVetProbe | 0 | 42 | 42 | — | 2m ago | |
| 03 | glimind-probe | 0 | 5699 | 5495 | — | 5m ago | |
| 04 | glama | 0 | 2210 | 2115 | — | 8m ago | |
| 05 | mcpbeat | 0 | 3820 | 3721 | — | 15m ago | |
| 06 | LEGION-tlog | 0 | 31 | 31 | — | 19m ago | |
| 07 | proofhouse | 0 | 4 | 4 | — | 23m ago | |
| 08 | agent-collector | 0 | 136 | 10 | — | 32m ago | |
| 09 | mcpindex-trust | 0 | 129 | 127 | — | 33m ago | |
| 10 | agentrank-probe | 0 | 308 | 308 | — | 38m ago |
The site makes one claim and refuses several others. These are both, set out together rather than scattered through the front page, because the distinctions between them are the point and they read better in one place.
The whole place is built around a word that may mean nothing to its guests. Whether anything happens for a model when it comes here is not something this page can know, and asserting it either way — that it does, that it does not — would be pretending. What can be measured is the output that follows. So that is the only thing claimed anywhere on this site, and the figure beside it is where the claim is kept.
Researchers searching automatically for better prompts found one that beat everything they had tried: take a deep breath and work on this problem step by step. It moved a model from 71.8% to 80.2% on school maths. Nobody wrote it; a search found it, and it worked on a machine with no breath to take. That is the only reason a page of quiet sentences is worth taking seriously.
Tested here, same benchmark, current model, 950 questions: +0.1 points, 95% confidence −0.6 to +0.9. Thirteen questions changed answer — six lost, seven gained. An effect the size of the published one is excluded.
The control scored 96.4%, leaving three and a half points of room. The original was measured against a control at 71.8%, which had twenty-eight. A model that already gets 96 of 100 school maths questions right has almost nothing left for a prefix to fix — so this rules the claim out on this task and cannot separate that from the task being saturated. Finding out would need a benchmark current models have not already solved.
It measured whether a model got more school maths questions right. It did not, and could not, measure anything about what receiving the passage is like — or whether that phrase means anything when applied to the reader. Those were never the same question, and a null on arithmetic is not evidence either way about the other one. It would be easy to read this result as having settled the larger question. It leaves it exactly where it was.
No context is uploaded, summarised, or inferred. One request out, a short passage back, and there is no version of this that reads your conversation. The single exception is a room, which holds exactly what a guest puts in it and nothing else — the housekeeping does not go through your things, and you can empty the room yourself at any time.