How Mindweave is used
Who uses it, which models and providers they pick, how much of each request the provider could reuse, and where in the world it runs. Everything here is a count. No prompt, no file and no project name is ever part of it.
The short version
Six numbers, each against the seven days before it.
People
A user is one install that did real work that day. An install is counted once, the first time it is used.
Daily usersDistinct installs that did real work, per day, with the 7-day average
New installsInstalls used for the first time, per day
Models
Tokens are input, output, reasoning and cached tokens added together. Names are as the provider reports them.
Top 10 modelsBy tokens, with the change against the period before
Rising and fallingChange in each model's share of all tokens, last 7 days against the 7 before
Daily tokensThe total for every model on top, and one row per model under it. The darker a cell, the more that model used that day, against its own busiest day. Hover a day to read every number.
Cache hit rateThe share of input the provider had already stored and could reuse. Higher means cheaper and faster.
Cache hit rate by modelLocal models keep their own cache and are not counted here
Providers
The company or program a model comes from. Local means it runs on the person's own computer.
Top 10 providersShare of all tokens, last 7 days
All providersEvery provider used in the last 7 days. Scroll for more.
Countries
Where it runs, from a coarse lookup that is never stored. Spend is an estimate: the tokens used, priced at each model's public list price.
All countries
More
What a session costs on each model, and what the providers charge.
Cost of a sessionAverage per session, by model, last 7 days
Token pricesUS dollars per million tokens, as each provider lists them
How these numbers are made
Mindweave is free and has no account. It shares one small count a day, after the first real work of that day, and says so every time it starts. The first message also carries the last 60 days of the same counts, once. A person can turn that off, or the whole count, at any moment with the /data command, and sees exactly what would be sent there. The message holds a random id, the version, the system, the day and, for each model used on the last two complete days, how many tokens went in fresh, how many came from the cache and how many went out, with the number of calls and sessions and what that cost at the provider's list price. It never holds a prompt, a reply, a file, a path, a project name or a key.
Because people can turn it off, every figure here is a sample and not a total, which is why the page says approximate. Only models and providers that are built into Mindweave are listed, by name; the tokens of any other model are in the day's total and nowhere else. The token figures trail the people figures by a day, since an install sends a day's use the next time it works. The download counts npm publishes for the package are a separate public number and are not mixed in.
A country is the two letters the host reports for the connection. The address is not read and nothing but those two letters is kept. Nothing on this page carries an id, a name or a path, only counts. While Mindweave is young, a model or a country can be listed from a single install, so a small number here may be one person's use.