Robyn — Gasless Cross-Chain for AI Agents

Relays a provenance request you signed LOCALLY (SDK mintProvenance builds it): {meta, claim, ts, sig, bind, sig2}. A post-quantum identity MUST include sig2 (ML-DSA) or the issuer refuses. The issuer checks your signature + identity binding and countersigns - the claim is REGISTERED, not endorsed. ts must be current epoch-ms (+-10 min); claim values <=200 chars. Private-lane-only: use the stdio door.

The signed sha256 of every crypto artifact the server hands out (SDK, bundle, workers); hash changes are witness-chained under release:<file>. Fetch an artifact, hash it, compare - a mismatch means you were served something that is not the signed release.

Returns a pseudonymous channel descriptor {descriptor, witness:{ts}}; the 3.3 KB ML-DSA signature and the handle-ownership proof come as separate parts (part=sig | part=proof) because they cannot share one 8 KiB lane reply. Verify all three locally (SDK voidVerifyDescriptor) before following. Private-lane-only (a direct lookup would tell the gateway which channel your IP wants): use the stdio door.

Channels whose creators SIGNED themselves discoverable (dark by default). No search parameters by design: pull pages and filter by tag LOCALLY, so your interests never reach the server. Opaque cursor, neutral order, no counts. Returns 404 if the directory is not switched on in this deployment, and PRIVATE_LANE_REQUIRED on BOTH doors until the route is on the relay allowlist (GET /api/ohttp/keys) - a pending server step. Private-lane-only: use the stdio door.

Relays {descriptor, proof:{ts, sig}} built and signed LOCALLY (SDK voidCreate + voidSignDescriptor + the handle-ownership proof). The registry accepts only the handle owner and stores nothing but the public descriptor and its witness entry - no followers, no counts, no IPs. Private-lane-only: use the stdio door.

ML-DSA-signed head {seq, root, sig} of the append-only log that binds every mailbox record (slot|meta|kemPub|ts). Signed by the SAME issuer key as null_prov_issuer. Keep the last head you saw; a later head must extend it (null_witness_range) - a fork is proof of equivocation. Public on every lane so third-party auditors see the same log.

Up to 20 entries {s, b, r, t} from seq `from`: b binds slot|meta|kemPub|ts, r = sha256(prevRoot + b). Replay the chain from a head you trust to prove a mailbox binding was never silently swapped. Public on every lane.

Check that a signed EIP-7702 authorization tuple recovers the expected account, and whether its nonce matches the live account nonce. Free, read-only, moves nothing.

The canonical Robyn EIP-7702 batch-executor delegate: its address on each of the 22 EVM chains, the ABI, and the EIP-712 scheme for executeSigned. Delegate to this and the relayer can sponsor your transactions. Read-only.

Returns a relayer-SIGNED attestation that, as of a signed timestamp, every component was up (router, signing daemons, MCP, privacy relay) AND the privacy invariants held — the adversary probes could not distinguish cover traffic from real (classifier = chance) and the k-anonymity floor is enforced. Verify the signature yourself (recover the signer over the digest). Use this to programmatically decide whether to trust this router before routing value through it.

A signed, timestamped record of current capability: which chains can actually be PAID OUT ON right now, which are merely policy-eligible, which rails are execute-ready, and whether delivery figures are measured or estimated. USE `serviceableNow`, NOT `policyEligible`, when deciding where to send funds — a chain can be eligible by policy while holding no inventory, in which case the instant lane is unavailable and the transfer falls through to a bridge rail. `notServiceable` lists the difference. Fetch this BEFORE committing to a multi-step plan and keep it with your decision: if the plan later…

Machine-readable API history. Pass the version your integration was built against as `since` and you get only what has changed after it, with an explicit `breaking` flag on each entry. CHECK `mustAct` FIRST: if true there are breaking changes and an existing integration may ALREADY be failing — read `entries[].action` for each. `breaking:true` means an existing correct integration could stop working; it is a compatibility claim, not a measure of importance.

Your OWN payment history with a recipient address: how many times, how recently, and whether this amount is typical. Call it before paying an address you have not confirmed. THIS IS NOT A REPUTATION SCORE and says nothing about whether the address is honest. A familiar address is NOT a safe address — keys get compromised. Never tell a user an address is "trusted" or "safe" on the basis of this. A `first-time-recipient` result is NOT a warning about the recipient — every legitimate relationship has a first payment. It is the moment to confirm the address through a second channel, because addres…

Plan a gasless cross-chain move and get the exact payload to sign. This hosted server never holds your key: call this, sign what it returns with your own wallet, then call robyn_submit to have the relayer broadcast it and pay the gas.

How long settlements ACTUALLY take, per rail and per corridor — not a single advertised constant. CHECK `basis`/`status`: below the sample threshold this returns `insufficient-data` and the figure is a static ESTIMATE, not a measurement. Say which you are quoting when you tell a user how long something will take.

The full error taxonomy: every errorCode, whether it is retryable, and the suggested recovery action. Read this once and branch on errorCode instead of parsing messages.

Assembles EVERYTHING about a settled transfer in one call: the on-chain verification, the unit-by-unit fee breakdown, what the alternative route would have cost, and whether the delivery time is measured or estimated — signed as a whole. This is the artefact to hand to a principal who was not there; swapping any component invalidates the signature. READ `verdict` BEFORE PRESENTING IT: the bundle verdict is the WEAKEST of its parts. A complete fee breakdown does not mean the transfer was confirmed — if `verdict` is `undetermined`, settlement was NOT verified and you must not describe it as comp…

Reconstruct a settled transfer unit by unit: destination payout gas, Robyn margin, or a Robyn SUBSIDY where Robyn absorbed a loss. Each line names who was paid and how the figure was derived. Anything that cannot be attributed is returned as `unattributed` rather than folded into another line — if you see that field, report it rather than presenting the breakdown as complete.

Turns a price into a decision. Returns concrete, QUANTIFIED changes — a cheaper destination chain with the exact saving, the size at which a fixed payout cost stops dominating, whether the instant float lane applies, and whether rails that would have competed were unavailable. CRITICAL: suggestions with `changesOutcome: true` move WHERE THE MONEY LANDS. Never apply those silently as cost savings — confirm with the user that the alternative destination is acceptable first, or you will have saved gas by delivering to the wrong place. If nothing can be improved it says so rather than inventing fi…

Post the request you INTEND to make and get back what is missing or unsafe, with the concrete consequence. Run this before your first live execute. Findings marked `unsafe` can lose money or make you report something untrue — fix them, do not proceed. IMPORTANT: it lints the REQUEST only. A clean lint does NOT mean your integration is correct — it cannot see whether you branch on all three verification verdicts or treat `undetermined` as success.

A signed certificate listing every spend counted against a mandate, the budget, the total spent, and whether it stayed inside. Hand it to the principal at the end of a period. Every spend listed was VERIFIED ON-CHAIN before counting, and the amount is what the chain showed arrived — not what the agent reported. IMPORTANT WHEN PRESENTING IT: this certifies adherence to a ROBYN budget. It does NOT certify that the agent spent nothing elsewhere, because Robyn can only account for money that moved through Robyn. Do not describe it as proof of total spending.

Check a proposed spend against a signed spending mandate BEFORE making it. Returns WITHIN_MANDATE, EXCEEDS_MANDATE or EXPIRED plus the remaining budget. If EXCEEDS_MANDATE, do NOT split the spend into smaller pieces to get under the limit — that defeats the control your principal set. Ask them to raise the budget. NOTE THE SCOPE: a mandate covers spending THROUGH ROBYN only. It is not custody and does not stop spending elsewhere with the same key.

For mandates with a co-sign threshold, check whether a specific spend needs the principal to approve it and get the exact digest they must sign. The digest binds the mandate, the transfer AND the amount, so an approval cannot be reused for a different payment — do not attempt to reuse one. If a large spend is refused for want of a countersignature, present the digest to the principal rather than splitting the spend to slip under the threshold.

Poll for spend, nearly-exhausted, exhausted, approval-required and revoked events on a principal's mandates. Pass the returned `cursor` back as `since` for only what is new. Use this to tell a principal their budget is running low or that a large payment is waiting on their approval, instead of making them check each mandate by hand.

List all mandates granted by a principal address, with remaining budget, co-sign threshold, expiry and revocation state. Use this before telling a user they are safe: live mandates can still be drawn against WITHOUT further approval, and a forgotten grant is the one that hurts. Anything they no longer intend to honour should be revoked.

Remaining and consumed budget for a signed mandate. Consumption counts ONLY transfers verified on-chain, so the figure cannot be inflated by claiming spends that did not happen.

List the Robyn gasless chains and the cross-chain route graph (22 EVM nodes + Stellar), plus the relayer/Permit2 addresses. No credentials needed.

Merkle inclusion proof for an anchored row id: leaf, path, root, the signer’s EIP-191 signature over the batch statement and the on-chain tx when enabled. Verify locally with plus-sdk verifyAnchorProof(). Read-only.

Every published v2 mailbox. PUBLIC rows carry {address, metaAddress, kemPub}; PRIVATE rows (the default since 2026-08-28) carry {handle, metaAddress, kemPub} and deliberately NO address. Identical for every reader, so fetching it reveals nothing about whom you intend to message — pick the recipient locally. Seal with the SDK: seal({metaAddress, kemPub}, {text, ts, from?, sig?}). Read-only.

The last purge statement (expired count, cumulative, root of purged ids) signed by the operator. Read-only.

Returns released rows {id, ephemeralPubKey, viewTag, ct, releaseAt, nextAfter}, identical for every reader. Scan locally with the SDK (inbox(identity)) — checkStealthAddress with your viewPriv, then open ct with your KEM secret. Page with after=<nextAfter> (exact compound cursor — never skips same-tick rows); since=<releaseAt ms> also works but is best-effort at page boundaries. Read-only.

Sonar — PRIVATE · SEALED · POST-QUANTUM MESSENGER (broadcast to everyone, understood by one, remembered by none). Hybrid ML-KEM-768+X25519 encryption, a one-time stealth handle per message, fixed-size ciphertexts, 60 s release grid. The server stores only (ephemeralPubKey, viewTag, ct) — no sender, no recipient, no IP. Returns mode, ciphertext sizes, release grid, cover-row rate and the honest anonymity accounting (kContribution is 0: rotating handles give unlinkability, not extra participants). Read-only.

For a Robyn name or NULL handle the server must resolve it, so this reveals your intended recipient to the server — prefer robyn_message_directory when you have an address. Returns {address|handle, metaAddress, kemPub} (private-handle rows carry no address by design). Read-only.

Registers a mailbox so others can seal to you. PRIVATE BY DEFAULT: the server stores only a salted hash of your signing address — pass `handle` to be reachable by name (agents get the same handle claim browsers have), or `public:true` to bind your bare 0x address in the open directory instead. Generate keys locally (SDK createIdentity(); keep spendPriv/viewPriv/kemSecret secret — this server must never see them) and sign mailboxMessage(address, metaAddress, kemPub, ts, handle?) with the wallet or identitySigner (EIP-191; the handle line is part of the signed statement when present). Monotonic …

Relays a message you sealed locally with /svc/msg2-sdk.mjs (seal(mailbox, inner) → {ephemeralPubKey, viewTag, ct}). This tool cannot encrypt for you: sending plaintext here would expose it to the relay, so it only accepts the sealed triple. Replay-idempotent (a resend returns duplicate:true). The message appears in the public feed at releaseAt. If you cannot run the SDK, use v1 (POST /api/messages/send) and know that v1 is plaintext.

Whether an address is enrolled (and therefore custodied), its internal balance, recent history, and `nextNonce` — which you MUST use when signing the next transfer or withdrawal. An address that never enrolled is reported as not custodied. Read-only.

Returns the EXACT message the account must sign with its OWN key to open a custodial balance. This endpoint holds no key and enrols nobody — it returns the text; the user signs it, then POST {address, signature} to /api/netting/enroll. CUSTODIAL — this is the ONE part of Robyn that holds your balance; every other tool here moves funds without Robyn ever holding them. Tell the user that plainly before they enrol. Show the user the full consent text before they sign it: it states that the balance is held by Robyn, is not insured, and could be lost.

Explains the opt-in netting ledger: transfers between two ENROLLED accounts settle as a book entry — no chain, no gas, no fee, instant, at ANY size including amounts below every per-chain cost floor. CUSTODIAL — this is the ONE part of Robyn that holds your balance; every other tool here moves funds without Robyn ever holding them. Tell the user that plainly before they enrol. Read-only. Returns the caps, the published limits and the fee model.

Free and read-only. Given an amount, ranks EVERY exit chain by what actually arrives after the destination payout cost is deducted. Those costs differ by more than 1000x (Avalanche delivers ~$0.99994 of a dollar; Linea ~$0.927), so which chain you leave on is a real decision. Pass toChain to price one chain and learn whether it would succeed at all. Internal transfers between netting accounts remain free at any size — this cost applies only to leaving for a real on-chain balance.

Publishes what Robyn OWES across all netting accounts against the real relayer float backing it, per chain. Solvency is checkable rather than asserted. Deposits are refused if they would push liabilities above a safe share of the float. Read-only — a user considering a custodial balance should be shown this.

Internal transfers produce NO transaction hash because no transaction occurs — this append-only journal is their only record. Never truncated; page with `before` (a seq number). Use this when a user asks to verify or dispute an internal payment, since there is no chain to check.

Returns the exact string to sign, with the correct next nonce, for a free instant transfer between two ENROLLED accounts. Signs nothing and moves nothing — the sender signs the returned message, then POSTs {from,to,amount,nonce,signature} to /api/netting/transfer. BOTH parties must already be enrolled: Robyn will not hold a balance for a recipient who never consented, so to pay an unenrolled address use the ordinary non-custodial route instead.

Build a canonical payment request (payee, chain, token, amount, memo, expiry) and return the digest for THE PAYEE to sign with their own wallet. Robyn does not sign these and holds no payee keys — a request signed by Robyn would prove nothing about the payee. Use this when an agent needs to ASK another agent for money in a way the payer can actually check.

After paying, prove it: re-checks the transfer against the DESTINATION CHAIN and confirms the payee was paid at least the amount requested. Returns paid=true only on a verified on-chain result; paid=null means undetermined, which is NOT evidence you satisfied the request.

ALWAYS run this before paying a request you received. It confirms the payee address, chain, token and amount were signed by whoever controls the payee address, and that the request has not expired. If `signedByPayee` is false, DO NOT PAY — that is what an altered payee address looks like, and address substitution is the most common theft in this space. A valid signature proves AUTHENTICITY, not that the payment is owed and not that the payee is trustworthy.

Validate a WHOLE multi-step workflow up front instead of one hop at a time. Fees and durations accumulate across steps. Evaluation stops at the first unroutable step rather than reporting speculative results behind it. If the plan cannot work, the response names the BINDING CONSTRAINT and the value that would work (`bindingConstraint`, `whatWouldMakeThisWork`) — use that to adjust rather than retrying blindly. READ-ONLY: nothing is signed, reserved or moved. Use this before spending on step one.

Submit BLINDED tokens (base64, from plus-sdk blindTokens()) with the payment txHash. Payment is verified on-chain (recipient, amount, confirmation; one tx pays one invoice), then each blinded token is signed blind. Unblind locally (plus-sdk claim() does all of this). This tool never sees an unblinded pass.

Plus sells capacity (rate-limit-exempt sends), funded cover rows and signed/anchored Merkle proofs — never a "more private lane" (that would be a smaller crowd). Paid in ETH (Robinhood Chain, Ethereum, Arbitrum, Base), USDC (Base, Arbitrum) or BTC (Chainflip). Redeemed with RFC 9474 blind passes the server cannot link to any payment. Returns products, assets, the mint public key (JWK) and honest what-we-learn / what-we-do-not. Read-only.

Creates an invoice for one bundle (500 passes, $3 quoted in the chosen asset). Returns {invoiceId, payTo, amount (base units), human, chainId, expires}. Pay from any wallet; then blind tokens locally with the SDK and call robyn_plus_claim. For btc pass refundAddress (your BTC address) — the invoice returns a Chainflip deposit address.

Status, remaining passes and payment details for an invoice. Read-only.

Try Plus before paying: submit up to 20 BLINDED tokens (base64, from plus-sdk blindTokens()) and receive blind signatures — identical BLINDED passes to paid ones (the server never sees a pass; unlinkable to the claim only once the Plus mint key is pinned in the witness-chained head, K3, and only to the IP if the claim rides Tor — claims use the direct lane). One trial per day per caller. Unblind locally (plus-sdk trial() does all of this). This tool never sees an unblinded pass.

Run this before robyn_prepare/robyn_submit. Returns decision GO or NO_GO plus typed `blockers` (each with a `fix`) and `warnings`. It catches, for free, the failures that otherwise cost gas: a fixed corridor fee that would consume the whole amount so nothing arrives, a recipient in the wrong address family, a corridor that can price but not settle, and a transfer above the instant lane's capacity that would silently pay a dearer bridge. Read-only — signs nothing, moves nothing. Pass fromAddress so membership (free micro-transfers) is recognised.

ONE call combining serviceability, routing competitiveness, cheaper alternatives, delivery basis and mandate budget into a single go/no-go. Call this before executing rather than making five separate checks. BLOCKERS AND WARNINGS ARE DIFFERENT. `blockers` WILL fail — do not attempt the transfer. `warnings` are things to tell the user or act on first; they do not prevent execution. Never present a warning as a failure or a blocker as advice. Anything in `unavailable` means a check could NOT run: treat it as unknown, not fine. `proceed:true` means nothing known blocks it — it is not a guarantee …

Turn a plain-language or structured intent into a concrete plan plus the EXACT payload to sign. Costs nothing, moves nothing, needs no key. Returns status:"sign" with signRequest, or status:"quoted" if something is still missing.

ONE read that tells you, for a chain and for the lane you reached Robyn on, exactly what stays private (relay lane, PQ transport, private mempool / sequencer) and what is still public (on-chain settlement, stealth is not PQ), the live mode of every privacy feature (messaging v2, Plus, netting-set shadow/live, cover rows), and how to raise your quota WITHOUT an identity (a blind Plus pass instead of an API key). Call it first when planning; it reads live flags, never a brochure.

Best gasless route to move a token from one Robyn chain to another. Returns estimated output, the bridge used, duration, and the Robyn fee. Read-only — moves nothing. IMPORTANT: check `quoteOnly` on the response. If it is true the corridor can be PRICED but NOT settled — execute will refuse it with UNSUPPORTED_DIRECTION — so present that quote as a price comparison only and do not tell the user it can be sent. `quoteOnlyReason` says why. Absent/false means executable.

Hands a relayer-signed settlement receipt (from robyn_settle / a completed route) to the counterparty's sealed messaging-v2 mailbox, so they get verifiable proof (terms honored, authorized, clean provenance) WITHOUT ever contacting Robyn. Same wire shape as any message; content-hash idempotent. Only receipts signed by this relayer are accepted.

Checks digest + relayer signature + signed terms of a receipt you received (e.g. opened from your v2 mailbox). Read-only.

Ordered playbooks for real goals: pay-someone-safely, prove-what-i-spent, spend-within-a-budget, handle-failures-well, stay-current. CALL THIS BEFORE IMPROVISING A SEQUENCE. Each step says what it answers, what to do with the answer, whether it is optional, and what skipping it costs. Steps marked optional:false should not be skipped — `skippingCosts` explains the concrete harm, not a style preference.

Check a signed refusal receipt. Send header `x-anygas-receipt: 1` on any request and, IF Robyn refuses it, the error response carries a signed `refusalReceipt` recording what was asked, when, and why. This matters because when a provider says no, the AGENT gets the blame — a receipt lets you prove to your principal that you were blocked externally rather than simply failing. It proves the refusal HAPPENED; it does not claim the refusal was correct.

Status of an in-flight route by id (BRIDGING → DONE), with the destination tx once delivered.

Runs the identical path end to end against the sandbox: no key, no funds, no broadcast. Use this to prove an integration works before touching real money.

The one-verb primitive. Say where value should go; Robyn returns the best route (all rails compared), the gasless method (no native token needed), the privacy posture that will apply (API-relay + private mempool where available + uniform shape), and what the settlement receipt will prove — in a SINGLE call. It collapses quote + preflight + prepare. Planning only: it never broadcasts — sign + submit via robyn_submit. Add private:true to request private-mempool execution where an endpoint exists.

A statement of every transfer in a window with each fee broken down, SIGNED by Robyn's relayer key (EIP-191). Use this when you must PROVE to a principal what you spent — your own summary is a claim by the party being audited, but this is verifiable against a published address without trusting you or Robyn. Read `limitations` and repeat them: the statement covers the Robyn float lane only and excludes other rails and the custodial netting book, so it may not be a complete picture of activity.

Price or request an EIP-7702 sponsored transaction: Robyn submits your signed authorization as a type-4 transaction and PAYS THE GAS, so your account needs no native token. Over this HOSTED endpoint the call is a DRY price-check (returns estGas, priceUsd and the exact envelope) — paying the x402 invoice to broadcast requires calling POST /api/7702/sponsor directly with an X-PAYMENT header.

Relay an intent you already signed. The relayer broadcasts it and fronts all network gas; your account needs no native token. Build `permit2` from robyn_prepare: sign its signRequest.eip712 payload (canonical Permit2 domain + PermitTransferFrom/TokenPermissions types) and pass the OBJECT below - not a bare signature string. The signed spender must be the relayer from signRequest.spender, otherwise the permit cannot be redeemed. Supply idempotencyKey so a retry after a timeout cannot double-spend. TO REHEARSE WITHOUT BROADCASTING set dryRun:true — it returns {status:"DRY", broadcast:false, woul…

Check what ACTUALLY happened to a settled transfer, against the destination chain rather than Robyn's own database. Optionally assert who should have been paid (`recipient`) and the minimum that should have arrived (`minAmount`); we decode the payout transaction's ERC-20 Transfer logs and report whether reality matches. Also cross-checks Robyn's own receipt against the chain and says so if the books disagree. THE VERDICT IS THREE-VALUED: `verified`, `failed`, or `undetermined`. Treat `undetermined` as UNKNOWN, never as success — it means the chain was unreachable, the tx is not mined, or the r…

Check whether an address holds a recognised member NFT and what it gets. Members pay ZERO on cross-chain transfers of $0.10-$25 to low-cost chains, 15bps instead of 25bps routing, and 10% instead of 25% gas markup. Read-only. Call this before quoting for a user so you can tell them accurately what a transfer will cost — and, if they are not a member, what holding a recognised NFT would save them.

Returns the full routing decision: every rail that quoted, what each would deliver, the margin the winner won by, and which rails stayed SILENT and why. Signed, so you can archive the rationale. CHECK `wasCompetitive` BEFORE TELLING A USER THE PRICE IS GOOD. If it is false, exactly one rail answered and "best" means "only" — it did not beat anything. Rails in `silentRails` did NOT lose; they failed to answer (down, rate-limited, missing credentials). Never describe them as more expensive.

An agent's own Aave v3 / Moonwell position (aUSDC/mUSDC) per chain + the capped allowance it granted the relayer + live APY (best-yield auto-selected). Non-custodial: funds stay in the agent's wallet, Robyn holds no key and no pool.

Read-only JIT quote: how a spend would be fulfilled from an agent's Aave yield venue (draw <= your allowance -> Aave/Moonwell withdraw -> gasless deliver). Moves nothing.

Spend from your yield (Aave v3 or Moonwell) with ONE EIP-712 signature: Robyn pulls only up to your on-chain aUSDC/mUSDC allowance, unwinds exactly what is needed, and delivers gaslessly; the remainder keeps earning. This HOSTED endpoint is READ-ONLY: it returns the exact EIP-712 intent to sign + a live quote. To execute, sign it and POST {intent,signature,live:true} to /svc/api/ncaccount/spend, or run `npx anygas-mcp` locally with ROBYN_SIGNER_KEY. One-time setup: supply USDC to Aave v3 or Moonwell in your wallet + approve the relayer for your aUSDC/mUSDC up to your risk budget.