A persistent, self-organizing terminal for agent CLIs.
Open a terminal, say what you want. Sessions run in real PTYs that survive
closing the app, every run's full output is kept — open any past session and
see exactly what it did — and saved commands become jobs that run on a
schedule, with claude, codex, gemini, or plain shell.
Internal tools
›"build us a purchase-order tracker: approvals, exports, a small dashboard"
An agent builds the tool in a conversation, then runs its routine parts on a schedule. Software you own, with no per-user license.
Server chores
›"set up nightly backups for these three servers and verify a restore weekly"
Set up once, then it runs every night — and once a week it proves the backup can really be restored.
Filings watch
›"every morning, pull the new 10-Qs for my tickers and note what changed"
A job checks for new filings every morning, downloads them, and writes a short note about what changed.
News watch
›"watch FT headlines and keep anything about chip export controls"
Checks every 15 minutes, slows down by itself when nothing happens, and speeds up again when news appears.
Weekly report
›"every Friday at 4, assemble the weekly KPI report from our exports"
Builds the report at the same time every week, checks its own output, and keeps proof that it ran.
Data migration
›"move these 2,000 invoices from the old system's exports into the new one"
An agent works through all the records, retries errors, and asks you only about the unusual ones.
Literature watch
›"collect new GLP-1 outcome papers weekly and file them by week"
Collects new papers every week, skips what you already have, and keeps everything on your machine.
Night runs
›"run the pipeline at 2am — if it breaks, fix it or ask me precisely"
If the nightly run fails, the cause is found and fixed automatically — or one clear question waits for you in the morning.
Persistent PTYs
Real terminal sessions that keep running when the app closes. Reattach
from anywhere; every run's full output is kept and openable any time.
Jobs
A command worth repeating becomes a job: cron schedule, run history,
output verification — and adaptive intervals that widen when nothing changes.
Any agent
Agent-agnostic by design: claude, codex,
gemini, or any CLI that runs in a terminal. Your subscriptions,
no API keys.
Self-healing
A failed run isn't an error message. It's diagnosed and repaired through
a deck of proven fixes — it only stops to ask you something precise.
Usage, watched
Per-agent sessions, active time, failures and subscription-window
estimates. You get warned before you hit a wall — never silence.
Local-first
One laptop or a quiet Mac mini in a closet. Your files, your keychain,
your history — nothing leaves the machine unless you say so.
Spend big once, rerun for ~nothing
exploreFULL AGENTWork the problem once — a capable model figures it out in a live session, on your subscription.
distillONE-TIMEKeep what worked — the working steps become a job with a verification check.
exploit~0 TOKENSRerun forever — scripts and small models do the routine; the big model returns only when something changes.
Roadmap
v1SHIPPEDThe persistent terminal — persistent sessions, jobs on cron, any agent CLI, usage watching.
v2IN PROGRESSDaily driver — runs unattended for days, full session/job control, signed & notarized app.
v3PLANNEDThe library — everything you collect in one place: full-text search, topics that assign themselves.
v4PLANNEDCloud & team — auto-sync, join by link, see and run machines anywhere.
v5PLANNEDSay what you want — typed intent routed to the right action; proven work distilled into cheap jobs.
[ @altoquant ]