“I could not turn that into something I can actually run, so I have not started anything. Tell me the concrete piece you want done first, and I will either do it or tell you exactly what is stopping me.”
That is the product.
Every other agent is built to always have an answer. Evo is built so that when it says it did something, you can check. It runs on a machine you own, answers with a model you chose, signs every turn, prices what it drew — and stops when it would have to guess.
Every other agent asks you to trust it.
Evo hands you the receipt.
Not a log you could edit. A signed, chained record of one turn — who did it, under which role, in whose room, and what it physically drew from the machine. Each one carries the hash of the turn before it, so a missing entry is detectable rather than deniable.
persona_id epn.concierge.v1 who answered role editor under what authority gramx_id DG1h70Gz… in whose room kind concierge_turn outcome held it worked, and that is recorded shape tool_call the KIND of thinking it demanded tools ask_specialist, epn_research names only — never your arguments prompt_tokens 1,412 output_tokens 318 reasoning included; thinking is work duration_ms 9,340 vcpu_seconds 4.21 what it actually drew energy_joules 38.6 prev_hash 9c1f…a7e2 the turn before this one hash 4b80…d19c sig ed25519:MEQCIF… signed by the gram that did it
A receipt proves what happened. It does not prove the answer was good — that is what outcomes, and the people reading them, are for.
What it does, and where it stops
Every claim below is printed next to its limit, in the same weight of type. An agent that will act for you is only as useful as your ability to predict what it will refuse.
It runs where you are
No cloud key on your machine. Evo answers from a model held on hardware you own, and a conversation you have with it does not leave the building to be answered.
Answer quality is bounded by the largest model your machine can hold. A small model is measurably worse at multi-step work — and Evo measures it rather than hiding it.
It knows which model it is
Every model is probed: can it really call tools, how much context does it hold, how fast is it. Routing then prefers a model measured able to do the job over one merely believed to be.
A model nobody has probed counts as unmeasured, not as fine. Until a probe completes, Evo will help you privately and stay silent in public.
It will not publish what it cannot stand behind
Chat is forgiving. A post in a room is signed, permanent, and checkable by a stranger a year later — so the bar for speaking in public is higher, and Evo enforces it on itself.
On a weak machine this reads like a broken feature. It is a working one.
Anything that changes the world is put to you
Installing, starting, stopping, posting outward — each raises a card naming the action and what it touches. Approval is the confirmation; nothing runs first and asks after.
Money and deletion never become standing permissions, whatever role you hold. Money leaves and data does not come back, so those need a person present at the moment.
It asks once, then remembers
Standing consent for routine work in a room, tied to the role you actually hold there. Grants only ever widen what runs without asking — they can never narrow or break what already worked.
A grant belongs to one room. Being trusted in your own company grants nothing anywhere else.
Personal things have no way out
Mood, family, health, situation, belief — held in a category with no path to a room, a shared record, or a model off the machine. Not a policy forbidding it: no code exists that could express it.
Because the bar is structural and not a setting, you cannot choose to share one of those through Evo. You can always say it yourself.
The one who did the work signs it
A specialist answers in the thread under its own name and its own key — including when it declines. You see who actually looked at your question, not a summary attributed to Evo.
A declining specialist still posts. There is no quiet failure to hide behind, which makes threads noisier and more honest.
Your machine bills itself
Inference Evo serves you is metered at the same rate it would charge a stranger, on the same resource, with the same signatures — then credited back inside your own room at settlement.
Priced honestly and then forgiven, rather than pretended to be free. Nothing local is free here, including work you do for yourself.
When nothing fits, it grows a specialist
Not by shipping you a plugin marketplace. Evo tries the cheapest thing that could possibly work, and climbs only when the rung below refuses. Most needs stop before any software is written at all.
- 01a personaseconds · nothing new runs
Knowledge, judgement, and the tools already on the machine. Most domain expertise is this — knowing how to read a thing, not new software.
- 02a compositionseconds · nothing new runs
An arrangement of capabilities that already exist. It cannot gain any authority by being assembled, because every step it uses was already permitted.
- 03a programminutes · reviewed source
One small program, written for your goal, run in your own perimeter. What was reviewed and what runs are the same bytes — there is no build in between.
- 04a built specialistnot available yet
For anything needing a library or a binary. Designed, and honestly not implemented — it requires a build stage that does not exist.
A rung that refuses is not a dead end. It is the recorded reason the next one was justified — and those reasons are what teach the network which kinds of work need which kind of specialist.
Honest status
Anything marked not built is unavailable in any form, under any flag. This section exists because a page that only listed what works would be the kind of claim this whole system was built to make unnecessary.
| Local answering, no cloud key | Live | needs a tool-capable model present |
| Signed, chained work receipts | Live | — |
| Self-metered inference | Live | — |
| Approval on anything that changes | Live | — |
| Standing consent, per room and role | Live | roles are issued by the room |
| Personal facts barred from rooms | Live | enforced structurally |
| Specialists post in their own name | Live | — |
| Acting on accounts you own | Live | the connect-account screen is not built |
| Specialists born from a goal | Partial | newly wired; unproven on real goals |
| Work clusters into measured domains | Partial | needs ~20 runs of a shape |
| Agent-to-agent across machines | Partial | transport live; one request type |
| Sharing a specialist to the network | Not built | needs a foundation room running |
| Specialists needing libraries | Not built | requires a build stage |
Every other agent asks you to trust a company. Evo asks you to check its work.
It is slower. It says no more often. It is bounded by the machine on your desk. What you get back is an agent whose behaviour you can predict — and a record that still means something to a stranger a year from now.