Describe an outcome in a sentence. OMNIRA plans it, shows you what it will cost, asks before it acts, runs it, and proves it worked — then keeps it running.
No card · no API keys · nothing is spent in practice mode
Worst case is above the $3 budget, so OMNIRA will stop and ask first.
A replay of real output from practice mode, where AI is simulated and nothing is charged
Not a chatbot and not a workflow builder. You state what should be true; OMNIRA works out how, and shows its working at every step.
You write a sentence. OMNIRA turns it into a typed goal — and asks when something is unclear instead of guessing.
Ordinary code, not a model, lays out the steps, picks who does each one and prices the whole thing.
You see what it will do, cost and touch. Nothing reaches outside without your say.
It runs, retries what is temporary, switches provider when one fails, and never repeats what already worked.
A run counts as a success only if its criteria hold against what was actually stored and sent.
Models propose.
Code decides.
Plans are assembled by deterministic code from a catalogue of typed steps, so the same request always yields the same plan — and a hostile web page cannot change it.
Risk tiers, budgets and approvals are rules. They are checked when the plan is made and again, step by step, as it runs.
If code can decide, code decides. A model is called when judgement is needed, and every call it was spared is counted.
OMNIRA starts narrow on purpose. Ask for something outside these and it will say so plainly — or ask what you mean — instead of pretending.
“Monitor these three sites every six hours. Tell me when something meaningful changes. Keep it under $3 a month.”
“Research solid-state batteries every morning and only tell me information I haven't already seen.”
“Send this JSON to my API whenever this webhook receives a new order, but never retry successful requests.”
Each run leaves an append-only ledger: what was done, who did it, which rule allowed it, how long it took and what it cost. The database itself refuses to edit or delete a line.
It is the audit trail, the debugger and the bill — the same record, so they can never disagree.
| 120 | EXECUTION_CREATED | plan v1 · estimated $0.000775 | |
| 126 | POLICY_ALLOWED | reading example.com was approved · tier 0 | |
| 130 | PROVIDER_SELECTED | safe HTTP client · score 0.74 | |
| 143 | TOOL_CALLED | news.ycombinator.com · 200 · 875 ms | $0.00000175 |
| 184 | AI_SKIPPED | a deterministic check showed AI was not needed | |
| 197 | BUDGET_CHECKED | worst case $0.0074 · within limits | |
| 207 | MODEL_CALLED | judge whether the change matters · 414 tokens in | $0.000714 |
| 231 | VERIFICATION_PASSED | 4 of 4 checks hold | |
| 232 | EXECUTION_SUCCEEDED | no important changes · 1.05 s | $0.000716 |
Software that acts for you has to be bounded before it is clever. Each of these is enforced in code and covered by tests.
Budgets per step, per run, per day and per month. Before a limit would be crossed, the run pauses and asks you.
Anything that acts on the outside world is recorded first. A delivery that succeeded is not repeated — even after a crash.
State lives in the database, not in memory. A worker can die mid-run and another finishes the job without redoing finished steps.
Web pages are data, never instructions. A model can describe a change; it cannot name a recipient, a target or an action.
Credentials are encrypted, referenced by name, resolved at the last moment and kept out of prompts, logs and the ledger.
Fetches are refused for private networks, local services and cloud metadata — including through redirects and DNS tricks.
Three pages fetched and compared by ordinary code. No model was called, so none was paid for.
Code filtered out the noise; a model judged the one real difference. That single call is nearly the whole cost.
The most likely monthly cost of checking every six hours. You see the full range, and the worst case, before you approve.
Figures are from runs in practice mode, where model prices are simulated and nothing is actually charged. Real prices come from a dated, versioned price list — a model without a price on that list is never used.
You will see the plan and its cost before anything runs