Nine agents, each sold on its own. Set up with the AI you already pay for.
The model proposes. The rules dispose. A person signs. Nothing is sent. Pick the agent you bought; the console walks a person at your company through the set-up in order, and your own AI (Claude, Grok, OpenAI, or another) can guide every step while the fixed rules verify every answer.
Step one: add your AI
Link the model you pay for, on your own key, sealed and never shown. You confirm by name that set-up guidance may use it. No model, no problem: the set-up works by hand too.
Step two: brief your AI
The console writes an installation brief from the agent's manifest: the steps, every question with its example and its check, the cards, the contracts, the cost and arm rules. Your AI reads it here, or you paste it into your own chat.
Every step verified
A rule checks each answer, a named person signs each card, a read probe proves each connector before it is kept, the cost is from figures you typed, the routine arms only on three signatures, and every number in a model's words is checked against its input.
Your tenant
Open an existing tenant
The key is sent only to your own function, at the address in config.json.
Create the tenant
The first tenant on a fresh install needs no install key; after that the install key set in your dashboard.
1 · Add your AI
The model you pay for, on your own key. It is sealed under your install's own secret and never returned. It is used for two things only: to guide the set-up (if you ask), and to write the words after the fixed rule has done its work. Linking it needs no signed card; a source does.
2 · Brief your AI
The installation brief for this agent, written from its manifest. Two ways to use it.
Let your AI guide you here
Your AI reads the brief and the state, asks the next question the rule names, and proposes an answer from what you say. Every reply is checked in code before it is shown; a proposed answer is stored only when you confirm it.
Or paste the brief into your own chat
Copy it into your Claude, Grok or other AI session; it carries no key, no tenant key and no prompt text. Then answer in the console.
3 · The interview
Ten fixed questions, one at a time, in your words. Each shows what a strong answer looks like. A vague answer gets a narrower question; nothing is invented from it.
…
4 · The card
Proposed by the rule from your answers; never approved by a model. Edit it, add to it, then a named person signs it. An edit clears an earlier signature.
Finish the interview, then propose the card.
5 · Link your connectors
Each a slot with a contract shown first. A send scope is refused. A read probe runs before the link is kept. The vendor is your choice, on your own licence.
6 · The cost, before the routine is armed
No vendor price is invented. For each linked connector that costs you lookups, type the unit cost on your plan or mark it included. A CSV, an ingest token and the queue cost nothing on the estimate. Not a quote from StoryC.
The estimate
7 · Arm the routine
Only when the three signatures exist. The database refuses otherwise.
The three signatures
The routine
Records for the next run
Paste rows as JSON (one object per line, or an array) in the shape the agent reads: . Or press "Load the specimen" to try the agent on synthetic records.
8 · Trust: the level, the prices a person authorised, the playbook, the ratings
The agent never raises its own level, never quotes a price nobody provided, and never cites advice outside the signed playbook. A person raises the level on evidence and can lower it in one click; any exception steps it down by the rule.
The autonomy ladder
Prices and discounts a named person authorised
The playbook, signed (optional)
Optional. Until a named person signs a playbook, the agent gives the facts in the data and no advice, which is safe. A playbook is the short list of facts the agent may cite, each with its source and the date it is next checked. Press the button to paste this agent's own research register, read it, change what you like, then sign.
Ratings and edits
The named person's queue
What the rule proposed, with a pointer behind every claim, and the model's words only when they passed every check. Status stays proposed until a named person decides.
Every proposed item
| Item | Kind | Score | Pointers | Status |
|---|
Receipts
Every run and every companion turn leaves a receipt: the prompt key and version, the model name or none, the hashes, the gates, the cap used, the reason anything was withheld.
| When | Run | Prompt | Model | Gates | Withheld |
|---|
Self test on this install: