v0.2 · self-host first · MIT

The agent that
grows with your work

Storms.CLI is a self-hosted autonomous runtime that lives on your server, remembers what it learns, and grows more capable the longer it runs. Bring your own keys, your own models — and never lose your work.

storms.cli — daemononline
06
subsystems
08
channels
MIT
license
100%
on your box
it gets better

An agent that compounds,
not one that forgets.

Every run writes episodes and distilled lessons to a local store. Nothing is uploaded, nothing expires, and nothing is lost when a model changes.

Day 1
Cold start

It reads your repo, your notes, your schedule. Every run is journaled from the first minute.

Week 1
Recall

Distilled notes start landing hits. Plans get shorter, tool calls fewer, results steadier.

Month 1
Routine

It schedules its own wakes, retries what failed, and stops asking questions it already answered.

Ongoing
Compounding

Memory keeps growing on your disk. Swap models any time — the learning stays with you.

the runtime

Six subsystems. One binary.

Nothing phones home. Every subsystem is a file you can read, a process you can kill, and a log you can diff.

Memory
Grows the longer it runs

Embedded vector + relational store on your own filesystem. Episodes, distilled notes, retrieval that never leaves the box.

Daemon
Resident and supervised

Crash-restarts, resumes its plan queue, appends every run to an on-disk journal you can read with cat.

Schedule
Unattended by default

Cron triggers plus reactive hooks. The agent decides when to revisit a task and requeues its own work.

Parallel
Delegates work

Isolated subagents with their own context, terminals and tool budgets — coordinated by one planner.

Sandbox
Real isolation

Every tool call runs behind an allowlist: filesystem scope, egress rules, hard wall-clock budget per step.

Replay
Deterministic

Every prompt, tool result and decision written as JSONL. Diff two runs, replay one exactly.

autonomy loop

It wakes itself, then improves itself.

1
Observe

A scheduled wake pulls fresh signals — files, repos, webhooks, your own notes — into a working context.

2
Deliberate

Specialist passes argue the read; a coordinator collapses them into one structured decision with a confidence band.

3
Act

Tools execute inside the sandbox under a hard time budget. Failures are caught, retried, or escalated.

4
Remember

The result is journaled and the distilled lesson lands in local memory, ready for the next wake.

run #8842 — replaydeterministic ✓
$ storms replay 8842 --diff 8830

wake     : cron(0 4 * * *)             budget  90s/step
context  : 6 signals · 2 memory hits · 1.8k tokens
tools    : repo.diff, note.recall, file.write
decision : ship  (conf 0.79)           prev: ship (0.71)
memory   : +1 note "changelog wants imperative verbs"
journal  : ~/.storms/journal/8842.jsonl  (14 events)
install

Two commands. One agent.

Bring your own keys, your own models, your own infrastructure. The agent belongs to you from minute zero — no account, no dashboard, no gateway in the middle.

Runs onLinux · macOS · Docker · Pi
Footprint~48 MB RSS idle
State~/.storms/{journal,memory,keys}
Keysyours — never transmitted
01install
$ curl -fsSL https://get.storms.cli | sh
02configure
$ storms.cli setup --sandbox strict
03run storms.cli
$ storms.cli daemon up --autonomous
the difference

Self-hosted, line by line.

criterionstorms.clihosted agent
Where it runsYour machineSomeone's cluster
Model keysYours, on diskUploaded
MemoryLocal store that growsVendor database
TelemetryNoneAlways on
Audit trailJSONL you ownDashboard export
Cost at idleElectricitySeat pricing

Own the loop.

Download Storms.CLI and boot it in under a minute. Self-hosted, replayable, MIT licensed — and it keeps getting better while you sleep.

runs autonomously on your infrastructure