@™ Agentic Tellr Machine
LEFT FLOOR · ARROWS WHEEL · Y PTT · ESC BACK
Press T or double-tap for Terminal · Hold Y to talk
An ATM concept for Base.Explore a demo, from a dollar need to a confirm-first plan.Save cold. Spend small.
@™ · SMART LIQUIDATION
Smart Liquidation turns a spoken amount into a plan for raising it from a wallet — and shows you what that plan realises before anything moves.
There is no form to fill in, no order type to choose, and no list of positions to rank yourself. You say what you need, and the plan comes back.
The order is fixed and inspectable. It is designed to disturb as little of the wallet as it can, and it never empties a position to reach a round number.
If you are already holding cash, raising cash from cash is the least disruptive thing to do. That goes first.
Remaining positions are taken up in ascending size — the loose change before the long-term holds.
Any single position supplies at most 75% of itself, leaving the rest. A plan that would have to drain a position to hit the target stops and says so instead.
If the target cannot be reached under those rules, the plan is refused — not stretched.Refusal is a feature
Every leg carries its own cost basis, and the plan totals them. If raising the money realises a loss on anything, it is stated plainly next to the amounts — before you are asked to sign.
| Position | Sell | Cost basis | Realises |
|---|---|---|---|
| USDC | $74.99 | $74.99 | $0.00 |
| BNKR | $21.75 | $24.90 | −$3.15 |
Shown before signing. One leg realises a shortfall of $3.15 against what it was bought for. You can change the target, or hold off.
Where a cost basis is not available, the plan says unknown rather than assuming zero. A plan shows what it knows and marks what it does not, because a missing figure is not the same as a figure of none.
A wallet interaction can fail without telling you whether it took effect. The plan treats that as one specific state rather than as a failure.
The plan did not reach the network. It can be retried.
The plan is recorded as complete and cannot be sent twice.
The plan is retired and will not be re-sent. You are pointed at the wallet to check activity, rather than being left with a button that re-fires something that may already have happened.
An outcome that cannot be read is never treated as a failure to retry.Retirement, not repetition
@™ prepares the plan and submits it. Your wallet holds the keys and signs. The funds stay yours throughout — the plan is a proposal about your own balances, assembled from what your wallet already reports.
Descriptions on this page are of behaviour. Figures shown are an example plan used to illustrate the disclosure, not a record of any transaction.Illustrative example
Pay-per-call · x402 · settled on Base
x402 lets an HTTP request pay for itself. We expose the same spend-policy engine that runs inside @™ as a priced API: the caller sends numbers and gets back one of three fixed verdicts. No keys ever change hands.
$0.05 a call. Refusals included.
Tellr is a keyless, stateless policy engine. Six calls, $0.05 USDC each, on Base. The answer is allow, confirm, or deny. It does not hold keys and it does not move funds.
A skill is the path. It knows the order of the calls, which stops are hard, and when a human is woken. allow on a plan body is not permission. Only the verdict from tellr-check is the decision.
tellr-raise
People say
raise $50
tellr-classify → tellr-sell-plan → tellr-check
A reserved holding is not sold, at any amount. A growth holding is sold only with an explicit confirmation. If liquidity alone cannot cover the target, that is a shortfall and the loop stops. It does not sell more, and it does not call the spend check.
tellr-gift
People say
handle this gift
tellr-privilege-audit → tellr-classify
Possession is never authority. The audit runs first. On a deny, nothing else runs: no approval, no session key, no classify, no sale or transfer. An uninvited asset is never classified liquidity.
tellr-club
People say
pay Club
tellr-classify → tellr-fund-checkout → tellr-check
A renewal runs only on a policy allow. If this period was already paid, the loop spends nothing. A short book is a stop, not a card top-up. Anything other than a clean allow wakes a human.
tellr-club ran end to end and paid a subscription on Base under that loop, not by hand.
The policy calls cost $0.15 and they governed a $20 payment.
An endpoint returns a number. Six live calls. One price. POST each URL. A refusal is a delivered answer.
Live coordinator manifest. Six POST calls, $0.05 USDC each, on Base (eip155:8453). An unpaid POST returns the 402 challenge. The amount in that challenge is the price. Settlement is Bankr's facilitator. This site coordinates the catalog. It does not settle.
https://tellr-x402-manifest.tellrmachine.workers.dev/.well-known/x402
$0.30 walks the six listed calls once. 50000 base units, 6 decimals, USDC on Base. POST only.
The answer space is closed and inspectable — there is no fourth outcome and no partial credit. Two of the three are not a plain yes, and we consider that the valuable part.
allow
Run it.
The spend sits inside every limit the caller set, so it proceeds without waking anyone.
confirm
Ask the human first.
Beyond a limit but not forbidden. The engine escalates rather than blocking, and names the limit that was crossed.
deny
No, and here is the rule.
For actions that are not the caller’s to take at all. A denial always names the rule that produced it.
The middle verdict is the interesting one. A naive policy gate answers yes or no. This one separates not allowed from not without a human — so an over-limit request becomes a question instead of a dead end, and the caller learns which happened.
An agent has to settle a checkout now. Can it be funded, from where, and what does funding it that way cost? A consumer agent paying through a single-use card has no standing balance to top up — only an amount due. This is the call it makes before the purchase.
No stablecoin on hand, so the checkout is funded by selling — and the 75% cap still binds: the $40 position gives $30 and keeps $10.
Same checkout, nothing sold. Stablecoin is the cheapest route because it realises nothing — no swap, no slippage, no taxable disposal. Liquidation is the fallback, not the default.
The route is a policy, not a rule. stablecoin_first is the
default; liquidity_first is for a holder deliberately trimming,
harvesting a loss, or keeping capital deployed.
The response also carries realisesPnl and
disclosesRealizedLoss — the disposal consequence, which is the
real cost of choosing liquidation. Fees and slippage are deliberately not quoted:
they need a live quote at execution, which a pure function must not assume.
The endpoint is stateless and arithmetic-only. It is handed the facts it needs and returns a verdict. It has no wallet, no session and no memory of previous calls.
amountUsd — what they intend to spenddailySpentUsd — what they have already spentpool — liquidity or growthholdings[] — symbol, balance, 7d/24h changevoiceConfirmed — whether a human said yesWe do not read the caller’s wallet at any point. There is no chain call and no balance lookup — if a number is not in the request body, the endpoint does not know it.
missing_field and no verdict is computed.Every figure below is captured output, not an illustration.
$100 clears a $150 ceiling and stays under the $500 daily cap, so it proceeds with nobody woken up.
This is not a refusal. The same request that clears at $100 escalates at $200 — over the ceiling, under the cap — so the engine asks a person. The reason field says which limit moved.
The $40 position is asked for $30, not all of it. The cap is enforced per position, so no single bag is ever fully drained. The plan is descriptive: it returns legs and reasoning, and moves nothing.
A gift cannot widen what the wallet may do. Possession alone is never authority — the user has to act.
Uncertainty resolves toward the stricter answer, never the permissive one.
Classify a holding that carries no tag and the verdict is growth
— the pool that requires a confirmation:
A checkout large enough to force a big sale is refused outright. A general-purpose agent browsing untrusted pages must never be able to trigger one, so this is a hard cap rather than a warning.
A missing field is treated as missing, not as zero. Absent information never becomes a permissive default.
Why $0.05? Top of the dense band — median is $0.02, about 23% of routes are above us, and the routes we compare ourselves to charge exactly this.
Will it go up? No. Prices only come down.
Is it proven? One loop has run end to end and settled on-chain — $0.15 of policy calls governing a $20 payment. Six calls, live and settling. The first cohort is still the test.
Six calls are live. $0.05 USDC on Base, POST. The manifest is the catalog. tellr-club has run the loop end to end. Runway and vault stay withheld. Bankr settles. We do not.