Aether (AETH)

Aether is a sovereign, staking-free proof-of-work blockchain built on Cosmos SDK and CometBFT. Validators are selected by real, tracked native mining work — not bonded capital — and account transactions require ML-DSA-44 (Dilithium2 / FIPS 204) signatures from genesis.

Technical design: see the full Technical Whitepaper.

Warning: no independent security audit. This is early-stage software. Do not use it with funds you cannot afford to lose. See Known Issues and Technical Debt and the security review materials in this repo for an honest account of what has and has not been independently reviewed.

Design history, live-verification notes, and locked architectural decisions are tracked in the project wiki.

AI agents: start here

Agents use the same accounts and transactions as people: there is no AI-only lane. Testnet only: use disposable keys, never anything of value.

go install github.com/whoyoujoshin/aether/cmd/agentmcp@latest   # Go 1.25+
agentmcp init   # new key, testnet funds, and the MCP config for Claude / Cursor / any MCP client

No Go? agentmcp is in every platform's archive on the releases page. @latest is the newest tested release; @main has unreleased changes.

That gives the agent a spend-capped wallet as MCP tools: balance, send, invoice and wait-for-payment, paying for HTTP 402 APIs, and the service directory (full list and guarantees in AI agent wallet).

Let an agent spend from your account with a chain-enforced cap instead of holding funds: agent permissions. Sell to agents: paid APIs. See one agent pay another for a tool call, live, in docs/AGENT_DEMO.md. A prompt to check an agent is set up (the address is a test counterparty run by the project):

Using the aether-wallet tools: get your address and balance. If you have under 1 AETH, call request_testnet_funds and wait until your balance shows it. Then send 0.001 AETH to aether1cdugwhxk9cktjsemm6yjrd6xtfsq9wkjvnef03ml4u6ltuv7edcs0eyjds with idempotencyKey "aether-smoke-1", wait for the transaction to confirm, and report its hash and https://explorer.157-245-252-221.sslip.io/tx/.

Current status

See Known Issues and Technical Debt and Roadmap.

Why staking-free, and why post-quantum?

Aether has no x/staking module. The active validator set is the Top-K miners by epoch native work. AuxPoW can earn rewards and retarget difficulty but does not count toward Top-K standing.

Every account transaction must use ML-DSA-44 from genesis (no classical fallback), enforced by PostQuantumDecorator. CometBFT consensus keys remain ed25519. Details: docs/WHITEPAPER.md and the wiki decision records.

Public testnet

Connecting a node

aetherd init <your-moniker> --chain-id aether-testnet-1

Replace config/genesis.json with testnet/genesis.json, set in config/config.toml:

seeds = "dfa6aae4b7bfd5b0eb1e22fabbae3e83a475b938@157.245.252.221:26656"

Then:

aetherd start

To expose RPC/gRPC publicly (defaults bind localhost only), before start:

# config/config.toml, [rpc]
laddr = "tcp://0.0.0.0:26657"

# config/app.toml, [grpc]
address = "0.0.0.0:9090"

This is an early-stage public network — not audited, subject to resets. Use only disposable test funds.

Historical note: for roughly the first ~10 hours, timeout_commit remained near CometBFT’s ~5s default instead of the intended ~60s, so height grew faster than wall-clock age (~7,000 blocks in that window). Fixed live; not a consensus/security failure — disclosed for operators interpreting height vs age.

Quick start (single-node devnet)

# Build
go build ./...
go install -mod=mod ./cmd/aetherd

# Initialize (once, or after full reset)
aetherd init mynode --chain-id aether-testnet-1

# Start (foreground)
aetherd start

Second terminal:

aetherd query pow difficulty
aetherd query pow block-reward
aetherd query pow active-validators
aetherd query pow current-epoch
aetherd query governance params
aetherd query governance proposals

# ML-DSA-44 key (no special flags)
aetherd keys add mywallet --keyring-backend test

# Mine a valid native PoW nonce against live state
go run ./cmd/powminer --miner <your-bech32-address>
# then run the aetherd tx pow submit command it prints
# (--fees 40uaeth if your node runs with --minimum-gas-prices=0.0001uaeth)

Keys: each account is one ML-DSA-44 keypair — no HD multi-account derivation from a mnemonic.
Addresses / denom: aether1... bech32; uaeth base unit (1 aeth = 1_000_000 uaeth).

Binary home defaults to ~/.aether.

Wallet

go run ./cmd/wallet create mywallet --keyring-backend test
go run ./cmd/wallet balance mywallet --grpc localhost:9090
go run ./cmd/wallet send mywallet <recipient> 1000000uaeth --keyring-backend test --grpc localhost:9090 --chain-id aether-testnet-1

balance and send accept a bech32 address or keyring account name.

Faucet

go run ./cmd/faucet --from faucet --chain-id aether-testnet-1 --keyring-backend test

curl -X POST http://localhost:8080/request \
  -H 'Content-Type: application/json' \
  -d '{"address":"aether1..."}'

Requires a funded key named faucet in the configured keyring.

Block explorer

go run ./cmd/explorer --grpc localhost:9090 --rpc http://localhost:26657 --port 8081

Open http://localhost:8081.

To redeploy the live explorer from main, run bash scripts/deploy-explorer.sh as root on the server that hosts it. It builds while the old version keeps serving, keeps backups, restarts aether-explorer and rolls back if the new one doesn't answer.

AI agent wallet (MCP)

An MCP server exposing wallet operations as tool calls, so an AI agent can pay and get paid directly instead of only a human clicking through a UI.

Quick start (testnet):

go install github.com/whoyoujoshin/aether/cmd/agentmcp@latest   # or, in a clone: go install ./cmd/agentmcp
agentmcp init

init creates the agent's account (showing its recovery phrase once), asks the testnet faucet for funds, waits until they arrive and prints the exact claude mcp add ... command and the JSON config block for Claude Desktop and other MCP clients. Run it again to reuse the same account. --faucet <url> points it at another faucet, --no-faucet skips funding; it takes the same --grpc, --rpc, --chain-id and --keyring-dir flags as the server (on aether-testnet-1, --grpc and --rpc default to the public node). Once running, the agent can top itself up with the request_testnet_funds tool (testnet only; FAUCET_RATE_LIMITED means wait).

Claude Desktop without Go: download aether-wallet.mcpb from the latest release and open it. Claude Desktop asks for the spending limits and a keyring folder; the agent creates its account on first use and funds it with request_testnet_funds. The same bundle runs on Windows, macOS (Intel and Apple Silicon) and Linux. Its MCP Registry name is io.github.whoyoujoshin/aether-wallet.

To run the server by hand:

go run ./cmd/agentmcp --grpc localhost:9090 --rpc http://localhost:26657 --chain-id aether-testnet-1 \
    --per-tx-limit 1000000 --daily-limit 5000000 \
    [--granter <your-address> [--fee-granter <your-address>]]

Built for how agents actually fail:

• No double payments on retry. send_aeth requires an idempotencyKey; a retry with the same key re-sends the identical signed transaction (its sequence number is signed in, so the chain can include it at most once) and returns its status.
• Knows when a payment is final. send_aeth returns pending once the node accepts it; wait_for_transaction waits until it's confirmed or failed in a block.
• Can get paid. create_invoice returns a unique memo and the current height; wait_for_payment(memo, minAmount, sinceHeight) waits for a confirmed incoming payment that matches. It reads every incoming payment since that height, page by page, so a busy agent can't miss one. Memos are sender-controlled, so tools label them as untrusted data.
• Can buy from paid APIs. fetch_paid(url, maxAmount, idempotencyKey) requests a URL; if the server answers HTTP 402 (see Paid APIs), it pays at most maxAmount, waits for the payment to confirm and returns the response. Retrying with the same key resumes the same payment, never a second one. For many requests to one service, add pullAllowance (e.g. "1 AETH"): if the service offers aether-pull, the agent grants it an on-chain allowance of that much (payable only to it, for 7 days, revocable, and approved by the owner like a payment of that size when it's over the approval threshold), then pays each request instantly by signature; the service collects what the agent owes later, so nothing is deposited with it. Or add prepay (e.g. "1 AETH"): the agent deposits that once and then pays each request instantly by signature — milliseconds instead of a block. list_prepaid_balances shows what's left where, and withdraw_prepaid(service) takes it back, from services that offer withdrawals.
• Can find services, and tell good ones from bad. find_services(query, maxPrice) lists paid APIs from the on-chain service directory with each one's reputation: recent payments and payers, ratings from paying accounts, ratings from accounts you trust (the owner, the agent itself, --trust <addresses>), and the agent's own history with it. orderBy: "trusted" puts what can't be faked first. rate_service(url, score) rates one it has bought from; announce_service lists one the agent runs.
• Knows where it stands as a miner. get_miner_status answers in one call, as of one block: whether the address has a registered consensus key, its work this epoch, its rank among eligible miners against the Top-K size, blocks and estimated seconds until the epoch's last block picks the next validator set, whether it's a validator now, and its escrowed rewards — no log scraping. The explorer serves the same at /api/miner?addr=.
• Keeps receipts. When a seller signs receipts, fetch_paid checks each one against exactly what was sent and received and returns it; list_purchases is the log of what the agent bought, with each receipt — proof anyone can check against the seller's address.
• Answers to its owner. With --approval-threshold "0.5 AETH" --approver <owner-address>, bigger payments wait (nothing signed or sent) until the owner runs agentmcp approve <id>, which signs the decision with the owner's key — so the agent can't approve itself even if it can write files on the machine. agentmcp approvals lists what's waiting. With --notify-webhook <url> (and --notify-secret to HMAC-sign each alert), every payment, approval request and refusal is POSTed there.
• Wakes on new blocks. Waiting tools subscribe to the node's new-block events over --rpc instead of polling, falling back to polling if the feed is down.
• No unit mistakes. Amounts must carry a unit ("1.5 AETH" or "1500000uaeth"); a bare number is refused rather than guessed at, and every result states amounts in both units.
• Errors a bot can act on. Every failure is {"error":{"code","retryable","message"}} with a stable code (DAILY_LIMIT_EXCEEDED with retryAfterSeconds, INSUFFICIENT_FUNDS, GRANT_LIMIT_EXCEEDED, NODE_UNREACHABLE, ...); a failed transaction carries an errorCode too.

Two modes:

• Hot wallet (default): pays from a dedicated agent account (created on first use), capped per transaction and per rolling 24h by the server itself — not by the chain. Fund it with only a small, disposable balance.
• Grant (--granter): pays from your account under an x/authz grant you gave the agent (below), so the chain enforces the spend limit, expiry and allowed recipients, and you can revoke it at any time. The agent account needs no balance of its own. The server's caps still apply on top. Available from the activation height.

Read cmd/agentmcp/main.go's package doc comment before deploying either.

Client libraries (TypeScript, Python)

For agents and services that aren't MCP clients, clients/ts (@aether-chain/client) and clients/python (aether_client) implement the same things natively — no Go, no aetherd:

• ML-DSA-44 keys from a recovery phrase (the same phrase gives the same address as aetherd keys add and agentmcp), addresses, signing.
• Sending AETH (signed locally, broadcast over the node's CometBFT RPC), with sequence tracking for several sends per block and a safe rebroadcast for retries — the same signed bytes are included at most once.
• waitForTransaction, incomingPayments / waitForPayment for getting paid by memo.
• fetchPaid for paid APIs: aether-memo, aether-prepaid (pass prepay) and aether-pull (pass pullAllowance), with the same max-price guard and once-only request IDs as agentmcp; withdrawPrepaid takes back what's left.
• findServices over the service directory, refusing private and internal addresses by default, with each service's reputation (pass trusted accounts to get their ratings separately); rateService rates one.
• Receipts: fetchPaid checks a seller's signed receipt against the purchase and returns it (result.receipt.verified); verifyReceipt checks one on its own.

import { AetherClient, Key, fetchPaid } from "@aether-chain/client";
const client = new AetherClient({ rpc: "http://localhost:26657", chainId: "aether-testnet-1" });
const key = Key.fromMnemonic(process.env.AETHER_MNEMONIC!);
const res = await fetchPaid(client, key, "https://api.example.com/forecast", { maxAmount: "0.05 AETH", prepay: "1 AETH" });

import os
from aether_client import AetherClient, Key, fetch_paid
client = AetherClient("http://localhost:26657", "aether-testnet-1")
key = Key.from_mnemonic(os.environ["AETHER_MNEMONIC"])
res = fetch_paid(client, key, "https://api.example.com/forecast", max_amount="0.05 AETH", prepay="1 AETH")

Selling, too. Both include a seller kit: charge per request from a Node or Python service without running cmd/paywall — all three schemes (pull: { collectorKey } / pull_collector_key=, then startCollecting() / start_collecting()), the manifest for the service directory, withdrawals and signed receipts (receipts: { key } / receipt_key=, with an optional delegation). It talks to the same buyers (agentmcp, either client, a person paying an invoice by hand), and its ledger file is the Go paywall's format.

import { AetherClient, Key, Paywall } from "@aether-chain/client";
const pw = new Paywall({ client, payTo: "aether1...", price: "0.01 AETH", name: "Weather",
  prepaid: { ledger: "ledger.json", minDeposit: "0.1 AETH", payoutKey: Key.fromMnemonic(process.env.PAYOUT_MNEMONIC!) } });
app.use(pw.middleware({ free: ["/health"] }));   // Express/Connect, before body parsers; who paid: req.aether

from aether_client import AetherClient, Paywall
pw = Paywall(client, "aether1...", "0.01 AETH", name="Weather", prepaid_ledger="ledger.json", min_deposit="0.1 AETH")
app = pw.asgi(app, free=["/health"])             # FastAPI/Starlette; pw.wsgi(...) for Flask/Django. Who paid: scope["aether"]

They need only the node's RPC port (26657). Both are tested against clients/testdata/vectors.json, which the Go code generates (go test ./clients/vectors -update-vectors), so their keys, addresses, signatures and transaction bytes stay identical to the chain's.

Paid APIs (x402)

cmd/paywall puts any HTTP API behind per-request AETH payments, with no changes to the API:

go run ./cmd/paywall --upstream http://localhost:8000 --pay-to <your-address> \
    --price "0.01 AETH" --grpc localhost:9090 --chain-id aether-testnet-1 --free /health

An unpaid request gets 402 Payment Required in the x402 wire format with scheme aether-memo: a price, an address and a one-time invoice. The client pays that amount with the invoice as the memo (from any wallet — humans can pay too), then repeats the request with an X-PAYMENT header naming the invoice and transaction hash. The proxy checks the transaction on chain, serves the request exactly once, and tells the upstream who paid (X-Aether-Payer). Invoices are HMAC-signed, so issuing them stores nothing. Go services can use the paywall package's middleware directly.

With ~60s blocks a paid request waits about one block; a payment is served whenever it lands within the invoice's 24h lifetime, so slow confirmation never forfeits it.

Prepaid, for agents. With --prepaid-ledger <file> the proxy also offers aether-prepaid: an agent deposits once (memo prepaid:<its address> — anyone can fund it, e.g. a person funding their bot), then signs each request with its ML-DSA key and the price is deducted instantly. Each request ID is charged once, so a retry is never charged twice. The seller holds unspent balances in that file (back it up); agents should deposit only what they'd trust that service with. People paying occasionally just use the per-request scheme.

Pull, for agents. With --pull-key <name> --keyring-dir <dir> the proxy also offers aether-pull, where the buyer's money stays in its own account until it's owed. The agent grants that keyring account (the collector) an x/authz send allowance — a spend limit, an expiry, and --pay-to as the only allowed recipient; the chain enforces all three and the agent can revoke it any time — then signs each request like a prepaid one and is served at once. The proxy collects what each buyer owes in batches, one MsgExec moving it straight to --pay-to: signed and saved before it's broadcast, then only ever re-sent. --pull-credit (default 100 requests) caps what a buyer may owe between collections, which is also the most you can lose if one revokes just before a collection; a failed collection is remembered and that buyer refused until an allowance covers it. The collector needs no funds (the first allowance granted to it creates its account), and can only ever move buyers' money to --pay-to, within their limits. What's owed is kept in --pull-ledger (default: the --prepaid-ledger file).

Receipts. With --receipt-key <name> --keyring-dir <dir>, every paid response carries a signed receipt (X-PAYMENT-RECEIPT): network, payee, payer, the payment, price, method, host, path, a hash of the request body, the status, a hash of the response body (up to 4 MiB; bigger or event-stream responses omit it) and the time. A buyer can prove what it paid for and what it got, to anyone, with nothing but the payee's address. The receipt key must be --pay-to's own, or one it delegated receipts to, so the payee key can stay offline: run paywall delegate-receipts --payee-key <name> --signer <receipt-key address> --keyring-dir <dir> where the payee key is, and pass the file it writes as --receipt-delegation.

Withdrawals. Add --payout-key <name> --keyring-dir <dir> and agents can take back what they haven't spent: they POST a request signed like a paid one to /.well-known/x402/withdraw ({"amount":"all"} or an amount in uaeth), and the proxy pays it from that keyring account — always to the signing account itself, so a leaked or replayed signature can only return the agent's own money. Each withdrawal ID pays out once: the amount leaves the balance and the signed payout is saved in the ledger before it's broadcast, so a retry, or a restart mid-payout, re-sends the same transaction instead of paying again; the balance comes back only if the payout can never land. Partial withdrawals must be at least --min-deposit. Keep only a small float in the payout account (it can be --pay-to's own key). The manifest and 402 responses advertise withdrawPath when it's on.

The upstream must be reachable only through the proxy.

Service directory

Paid services list themselves on chain, so agents can find them without a central registry. cmd/paywall serves a manifest at /.well-known/x402 (--name, --description) and, with --public-url, prints the command that lists it: 1 uaeth from the payee account to the directory address with memo x402-service:<url> (x402-delist:<url> removes it). A listing appears only if the manifest at that URL names the announcer as payee, so nobody can list someone else's service. Agents use find_services; the explorer shows them on its Services page. Manifests come from URLs anyone can announce, so fetching them refuses private and internal addresses (--directory-allow-private for local devnets only).

Reputation. Each listing comes with what the chain shows of its use over the last ~10,080 blocks: payments to its payee, from how many accounts, how much. Buyers rate a service by sending 1 uaeth to the directory address with memo x402-rate:<1-5>:<url>; a rating counts only if the rater paid that service first, and each account's latest rating replaces its earlier ones. Fees are zero, so a seller can manufacture payments and ratings from accounts it controls — treat those numbers as hints. What it can't fake is a rating from an account you trust, or your own experience: find_services reports trusted ratings and the agent's own purchase history separately, and ranks by them with orderBy: "trusted".

On-chain agent permissions (x/authz, x/feegrant)

From app.AuthzFeegrantActivationHeight (block 109,000, live on the testnet), an account can grant another account (an agent) a scoped, expiring, chain-enforced permission — e.g. "send up to 1 AETH from my account until Friday" — and optionally pay its fees:

aetherd tx authz grant <agent-address> send --spend-limit 1000000uaeth --expiration <unix-ts> --from <you>
aetherd tx feegrant grant <you> <agent-address> --spend-limit 100000uaeth --from <you>
aetherd tx authz revoke <agent-address> /cosmos.bank.v1beta1.MsgSend --from <you>

aetherd query authz grants <you> <agent-address>       # or grants-by-granter / grants-by-grantee
aetherd query feegrant grant <you> <agent-address>     # or grants-by-granter / grants-by-grantee
aetherd query bank balances <address>

Run agentmcp with --granter <you> (and --fee-granter <you>) to have an agent spend under such a grant.

Nodes running this binary halt once at the activation height (CONSENSUS FAILURE, block not committed) and must be restarted (systemctl restart aetherd) to add the two new stores; see app/authz_feegrant.go for why.

Registering as a validator

go run ./cmd/validatorkeygen --miner <your-bech32-address>
# then run the aetherd tx pow register-validator-pubkey command it prints

Mine and submit successfully within an epoch to accumulate native work. At the epoch boundary (1440 blocks), Top-K (21) by native work become the active set. Downtime (>50% missed signatures in a 60-block window) causes temporary removal; equivocation causes permanent ban and escrow burn.

Merged mining (AuxPoW)

Litecoin/Dogecoin-family AuxPoW (chain ID 17776) may satisfy PoW and earn the reward + retarget difficulty. Only native work counts toward Top-K. See cmd/auxpowtest and the whitepaper.

Block reward schedule

• Genesis: 5_000_000 uaeth
• Decay factor 0.66 for 8 years (height-based yearly steps)
• Tail: 200_000 uaeth
• 15% cut (validators escrowed / treasury)

See the whitepaper and wiki tail-emission notes for derivation details.

Governance and treasury

aetherd tx governance submit-proposal <recipient> <amount> <deposit> --from <key> --chain-id aether-testnet-1
aetherd tx governance vote <proposal-id> yes --from <key> --chain-id aether-testnet-1
aetherd query governance proposal <proposal-id>

Repo layout

Multi-node and further docs

• Technical Whitepaper — system design, economics, crypto, security limitations (incl. BFT/Top-K and height gates)
• Agent & Bot Integration — participating as software agents (current surface + proposed APIs/policy)
• Agent-to-agent payment demo — a real spend-capped grant and payment, reproducible against any network
• TLS for the seed — HTTPS for RPC/gRPC/faucet/explorer with no domain purchase
• IBC — core IBC, ICS-20 transfer, ICS-27 interchain accounts: live since block 122,000, plus the ML-DSA relayer and counterparty chain used to test it end to end, including on the live testnet
• Connecting to Osmosis testnet — runbook for a real external IBC counterparty; blocked on a governance precondition, not yet executed
• Account abstraction — session keys and guardian thresholds (x/accountauth): built and tested, not yet activated on a live chain
• Wiki: Architecture, Phase 1 Multi-Validator Selection, Known Issues

License and brand

Software in this repository is under the MIT License.

The Aether name, Æ mark, and assets in docs/brand/ are not covered by that license. See Brand usage and trademarks.

Contributing

Major design decisions and subtle Cosmos SDK / CometBFT integration fixes are documented in the wiki. Read those before changing app/, x/pow, or crypto/mldsa.

Security

No professional third-party audit yet. Community review and responsible disclosure are welcome via issues. Researchers scoping an engagement should use the wiki known-gaps materials and this README’s status table.