1.03
0.16
basin
$ pip install basin-cliopen source · an optimizer, a neural network and your ai agent · in your terminal and in claude code
scroll. the water follows. click the axolotl.
00 · filmtwenty-six seconds. it loops. click to pause.
01 · thirty seconds
two commands.
then one line a day.
then one line a day.
$ basin add "sleep before 00:30" $ basin step
every number below is computed from the files. no account, no sync, no server. basin demo generates the same numbers you see here.
0days on the slope
0.00loss today
0%water level
0tomorrow, pages
02 · one stepdrag the days. the same engine renders the card.
day 12
loss 0.00 → 0.00
slope gentle
03 · plateaukeep scrolling. the lines flatten.
descending
a plateau is not a failure. the water is still there. basin says so in one line and does not ask you how you feel.
04 · what it refuses
it will not take
a wish.
a wish.
the rules below are the ones the program runs. type something. anything without a number and a direction is refused, politely, once.
try: basin add "be better"
or: basin add "11 pages, more"
or: basin target
or: basin add "11 pages, more"
or: basin target
$
05 · born in a blockrobinhood chain · read-only · connecting
latest block
—
hash
every familiar, every roll, every card carries the block it was born in. the hash is the only randomness basin uses.
roll · the water decides
coolwarm
06 · your agent
claude code
says hello with
the step.
says hello with
the step.
your ai agent is the interface: mcp tools, a skill, three hooks. install the plugin and every session opens with your slope. the agent turns a sentence into a log line, reads the step, feeds the familiar with every tool call. it cannot see the target, because there is none stored. basin itself calls no language model and needs no api key.
$ claude plugin install basin $ codex mcp add basin -- basin mcp
07 · the familiara cellular automaton, live. click it to burn a hole.
your agent has a familiar. it grows from what the agent does. you cannot draw it, only feed it.
species
—
—
state
hatching
hatching
cells
0 / 0
0 / 0
born in block
—
—
a neural cellular automaton: a neural network with about six thousand weights, trained once to grow the axolotl from one cell and to heal damage. this page runs a simplified rule until the trained weights arrive; the cli runs the real one in pure python.
sixteen species · one per birth-block hash mod 16click one to hatch it
where the ai is · three places, none of them a chatbot
the step is gradient descent.loss, gradient, an adaptive learning rate with warmup and a cap. the arithmetic that trains a neural network, pointed at your week instead of at a model. one optimizer iteration a day.
the familiar is a neural network.a neural cellular automaton, ~6 000 weights, 16 channels, 64 hidden units, trained once with pytorch. ships as weights.json, runs in pure python, no torch at runtime.
your ai agent is the interface.claude code, codex or cursor talk to basin over mcp. the agent logs your day from a sentence, says hello with the step and feeds the familiar. basin itself calls no llm and needs no api key.
08 · the tokenthe program does not touch it.
networkRobinhood Chain
ticker—
contract—
swap—
chart—
explorer—
the chart mounts here by itself when the contract address appears.
a community token. no promises. the code does not know it exists.
a community token. no promises. the code does not know it exists.
09 · honest limits
basin only measures what you type in, plus the few sources it can read. it does not read your calendar, your watch or your bank.
the slope is an estimate from few points. the first three days it prints nothing but the next point.
the familiar is a toy. it will not help you refactor. it grows either way.
roll uses a public block hash. anyone can predict it a few seconds ahead. do not roll for money.
the token is not part of the software and the software will never require it.
the axolotl is not real. real ones are critically endangered.