Your agents decide when. You decide whether.
ULTIMA lets automation run on its own without letting it spend, send, or break anything on its own. Every consequential action stops for a human and waits — enforced in deterministic code, not by asking a model to behave.
No card. A perpetual licence, free. Runs on your machine.
This is a real session, captured from ULTIMA 0.1.0 against a live model and a live Telegram bot. The approval genuinely stopped; the disk figure is the machine it was recorded on; the packet is what you actually receive. Playback is time-compressed — the run really took the 28.9 seconds the cost line reports, most of it waiting on the model — and long blocks are trimmed where marked. Nothing is reworded and no line is invented.
Three different kinds of work, in order: reading the machine and writing a file — where the recovery ladder replans a step that had nothing to deliver; searching the web — which stops, because anything leaving the machine is L5; and sending a message — with the decision packet you actually get.
What it actually enforces
Outward actions stop
Anything that sends, transacts, actuates, can't be undone, is high-risk, or costs more than $0 is forced to human authority. The classification is code, so it can't vary run to run.
Failures don't loop
A "repair" identical to what already failed is refused before it runs. Attempt caps live in the database, so restarting doesn't reset them.
Half-sent stays stopped
An action that already reached outside the machine gets no automatic retry at all. A send that failed may have half-succeeded, and the kernel can't tell. That's your call.
Bounded by the clock and the budget
Daily ceilings on runs, compute and notifications — re-checked before every run, not once at the start. Quiet hours. A kill switch.
Everything is replayable
Every authority check, denial, approval and standing authority is
recorded. ultima replay <id> reconstructs any run.
Nothing phones home
Python and SQLite, no third-party runtime dependencies, no licence check-in, no activation. You bring your own model.
What it is good at
Every item below is something it was watched doing, not a category we think it should suit.
Recurring checks on machines you own
Measure, compare against last time, report only what changed. Readings are verified against a prediction before they are believed, so a broken probe reads as a mismatch rather than as good news.
Jobs where gathering is safe and sending is not
This is the shape it fits best. Collect, compute and write locally without asking; stop dead at the step that emails, posts, pays or notifies. The expensive half runs unattended, the dangerous half waits.
Automation you would not otherwise leave alone
Daily ceilings on runs and compute, quiet hours in your local time, a kill switch, and a cap on repair attempts that lives in the database — so restarting the process does not reset it.
Flaky integrations
A failure is diagnosed from what the error actually says and answered with the cheapest plausible repair. A "repair" identical to what just failed is refused before it runs. Anything that already reached outside gets no automatic retry at all.
Work that has to be explainable afterwards
Every authority check, denial, approval and standing authority is
recorded. ultima replay <id> reconstructs a run: what
it predicted, what it observed, and where the two disagreed.
Anywhere the data cannot leave
Python and SQLite, no third-party runtime dependencies, no licence check-in, no activation. It reaches the network only through adapters you grant, and each of those is an outward action that stops for you.
Where it fits badly, so you do not find out later: it is Linux only — it uses a per-user systemd timer and POSIX file locking, so there is no Windows build today. A run takes tens of seconds, most of it waiting on your model, so it is wrong for anything interactive. It is single-operator: there are no teams, roles or shared queues. And it does not write your software or run your business — it governs automation you already have.
What it is not
ULTIMA is alpha software and a deliberate subset: 25 of its 60 specified sections are built, 27 are partial, and 8 are not built. It does not run your business, generate your agents, or model your workflows — it governs automation you already have. The per-section ledger ships with the product. Evaluate it in an isolated environment before giving it access to real systems.
Free for the first 100
A full perpetual licence, at no cost, for the first hundred people who ask. Not a trial, not a cut-down build — the same signed release that is sold, and it never expires.
- No card. None is asked for, and nothing is stored that could be charged later
- A perpetual licence, twelve months of maintenance, email support
- We will ask once, after a week, what you think — and that is all it is: a question
- Say no to the review and you keep the licence anyway. There is no catch and no bill
Once the hundred are gone the price is $200, once, per Host.
One price, paid once
- Perpetual. The licence never expires and there is no recurring charge
- A second non-production Host for staging or standby, free
- Twelve months of maintenance included — every version released in that period
- Best-effort email support, answered by the person who wrote it
- No licence check-in and no activation, so nothing we do can switch off software you already have
3–9 Hosts $160 each · 10+ $130 each. Prices exclude tax. Maintenance renewal after the first year is optional, $49 per Host per year — the licence keeps working either way.
Evaluate it first
30 days, one non-production Host, no card. It is the same signed build that is sold — not a crippled demo. Email ultima@leadrescue.us and say what you want to point it at.
# verify before you install — this is the whole point sha256sum -c SHA256SUMS gpg --import ultima-signing-key.asc gpg --verify ultima_kernel-0.1.0-py3-none-any.whl.asc \ ultima_kernel-0.1.0-py3-none-any.whl gpg: Good signature from "Lead Rescue LLC"
Confirm the key fingerprint through a channel the download does not control before trusting it.
Questions
What happens if you stop operating?
You keep running it. ULTIMA ships as readable source, runs entirely on your infrastructure, and performs no licence check-in or activation of any kind. Nothing we do can switch off software already delivered to you — that is a term of the agreement, not a promise.
What happens after the first twelve months?
Nothing, unless you want it to. The licence is perpetual and keeps running. What ends is the maintenance term — new versions and support. Renewal is optional at $49 per Host per year, and you can pick it up again later.
Does it need an internet connection or an API key?
No. It needs Linux, Python 3.10+, and a model you supply — a local Ollama install is enough. Optional adapters can reach out, and every one of them is an outward action that stops for you.
Is it audited?
No. It has a large deterministic test suite covering the safety properties, and it reproduces byte-for-byte across Python 3.10 to 3.14. That is not the same as an external security audit and we don't claim it is.
Can I modify it?
Yes, within the agreement's terms — it ships as readable source. The licence sets out what modification permits and what it does not.