Agent kit · operate the launchpad with an AI

Hand the launchpad to your agent.

HyperPad ships a published, machine-readable skill so an AI agent — Claude, or any wallet-bearing bot — can operate the EVM rail on your behalf: find a sale, buy into it, claim vested tokens, refund a failed sale, watch for triggers, and (if you're a project) launch a token. You delegate the task; the agent reads the skill, does the on-chain work with its own signer, and reports back with verifiable tx hashes and amounts. Everything below the discover layer is a direct contract call — no hosted service to trust between the agent and the Sale contract.

01What the kit is

A SKILL.md entrypoint (a Claude Agent skill, name: hyperpad) plus five task skills, a full machine reference, and the contract ABIs & addresses. An agent points at the hosted skill, reads it, and then operates entirely through public RPC + the launchpad contracts. The kit is wallet-agnostic: the agent brings its own signer (an imported key, a managed CDP/Privy wallet, a Bankr-style bot). Read-only flows need only an RPC; buys/claims/refunds/launches are direct, signed contract calls.

02The five skills

Each skill is a focused, self-contained playbook. Three are read-only or position management; two spend funds.

• discover.md read-only · no funds

  Find and vet sales. Lists every sale from the factory, decodes getSaleData() + phases, computes human-readable price & progress, and returns a verdict — price, your max buy, window, vesting, whether LP is locked, and red flags (no LP, whitelist-gated, already ended).

• participate.md spends funds

  Buy into a sale. Resolves the pricing mode (FIXED or ORACLE_USD) and active phase, sizes the buy against the budget / per-wallet max / cap room, handles the whitelist Merkle proof, simulates, then sends buy (native or ERC-20 with an approval first) and reports the credited claim.

• claim-refund.md manages position

  After a sale settles, decide and execute. On success it claim()s vested tokens on the TGE + cliff + linear schedule (repeating as more unlocks); on failure it refund()s the full contribution. Enforces claim-XOR-refund and flags escaped sales (no recourse) honestly.

• launch.md spends funds + commits tokenomics

  Launch a token as a project. Gathers token/sale/LP/vesting/anti-snipe parameters, builds the SaleConfig + TokenParams structs (exact field order), checks supply covers sale + LP via tokensRequired(), simulates, and sends createSaleWithNewToken — confirming every parameter first, because terms are immutable after launch.

• monitor.md the "do it while I'm busy" loop

  Watch a sale and act on triggers: buy at open, top up before sellout, claim each unlock, refund if it fails, alert on settlement. Polls getSaleData() (or subscribes to events), then hands off to participate / claim-refund — with budget and idempotency re-checked every iteration.

03The safety model

Every spending action is bound by the rules baked into SKILL.md. An agent that ignores them is off-script.

• Budget is a hard ceiling.

  The user's stated spend is never exceeded — the buy size is clamped to min(budget, cap room, per-wallet room), and if it has to clip below what you asked, it tells you why.

• Simulate before send.

  Every state-changing call is eth_call'd from your address first; on a revert it surfaces the reason and stops, never blindly broadcasting.

• Re-read state at action time.

  Sales change with time and blocks, so getSaleData() is read in the same minute it acts; the oracle feed is re-read immediately before an oracle-priced buy.

• Verify addresses against the chain.

  A sale address must come from factory.allSales(i) or a SaleCreated event — never a raw link — and the token's decimals/symbol must match your intent.

• Decimals are explicit.

  Quote amounts use the quote token's decimals, token amounts use the sale token's — never assumed 18. (USDC 6 vs token 18 is the #1 agent mistake.)

• Idempotent spends.

  Before any retry it re-reads contributed / claimed, so a missed receipt can never cause a double buy or double claim.

04Use it — point an agent at the hosted skill

The skill is served from this site. Give your agent the URL and the task; it fetches SKILL.md, follows the links to the task skill and reference it needs, and operates with its own wallet.

/agent/SKILL.md ← the entrypoint to hand over

# Point your agent at the hosted skill, then delegate in plain language:

  Read the HyperPad skill at https://<this-site>/agent/SKILL.md, then:
  "Find the ACME sale, vet it, and if LP is locked buy 200 HYPE at open
   — never spend more than 200. Claim my tokens as they vest."

# The agent will:
  discover.md  → list sales from factory 0x0F3e…48D8, find ACME, return a verdict
  monitor.md   → wait for state Pending→Active, then
  participate.md→ simulate + buy ≤ 200 HYPE, report tx hash + tokens credited
  claim-refund.md→ claim each unlock until fully vested

HyperPad · agent kit · source: SKILL.md · skills/ · reference/ · unaudited — verify everything on-chain · back to the app