NFT Collector Copilot

An AI agent for NFT collectors. Watches OpenSea, scores every listing across five signals (price, volume, depth, holders, momentum), and executes Seaport trades through a Privy server wallet with a TEE-enforced per-tx cap and an off-machine owner key the agent cannot reach. You approve; it acts. The wallet's float is your real aggregate ceiling — you fund and replenish it on your own cadence.

The killer feature: Whale-Cross Alerts

Point it at a list of wallets you respect (vitalik.eth, your favorite PFP whale, whoever) and your watchlist. When a tracked whale buys, mints, or lists into any collection — whether it's on your watchlist or not — you get a one-line alert within a heartbeat, with a one-click follow gated by your Privy spend cap. Few others ship this — it takes the OpenSea events feed, a TEE-enforced wallet, and a disciplined scoring rubric in the same box.

What it does

• Watchlist floor + volume tracking with per-slug alert thresholds and sell-side flip triggers.
• Conviction Score before any recommendation — price vs 7d median, 7d/30d volume ratio, listing depth, holder concentration, momentum.
• Pre-Buy Gate — wash-trade check, thin-market refusal, gas economics, full-cost (with fees) disclosure, balance & buffer check, policy-fit check.
• Seaport buys + sells via @opensea/cli + the skill's fulfillment scripts, signed by your Privy wallet.
• Drop Radar — upcoming/featured drops cross-referenced against your taste model.
• Taste learning — every buy/pass/ask updates a structured taste.json so recommendations get sharper over time.
• Spend-cap safety. Per-tx cap enforced in Privy's TEE; owner key off-machine so the agent can't rewrite the policy; aggregate spend bounded by the wallet float you fund. Three independent layers — one in the user's hands, two in Privy's, none the agent can lift.

Supported chains: see workspace/TOOLS.md. Mainnets only by default.

What's bundled

• @opensea/cli — installed globally at build time.
• opensea-marketplace — attached via manifest.json → skills and mounted at skills/opensea/. SKILL.md + reference docs + shell scripts for Seaport, swaps, stream events, wallet setup, and policy templates. Pinata pulls the latest published version at deploy time.

Secrets you'll need

Only two are required up front. The rest the agent provisions for you on first run.

Variable	Required	How to get it
PRIVY_APP_ID	yes	dashboard.privy.io → create app
PRIVY_APP_SECRET	yes	Same page as App ID. Pair with an owner_id registered on the wallet — bootstrap walks you through it.
OPENSEA_API_KEY	optional	If unset, the agent fetches a free-tier instant key on first run and attaches it via the bundled Pinata Platform skill. Provide your own paid key here only if you need higher rate limits.
PRIVY_WALLET_ID	optional	If unset, the agent creates a new server wallet via opensea wallet create once Privy app creds are present. Provide one only to reuse an existing wallet.
PRIVY_AUTH_SIGNING_KEY	optional	Auto-generated by the agent on first run via opensea wallet generate-auth-key. The matching public key gets registered as the wallet's additional_signer; the owner key stays off-machine on your host.

Full walkthrough: skills/opensea/references/wallet-setup.md. Policy templates: skills/opensea/references/wallet-policies.md. Hot-wallet float pattern: skills/opensea/references/wallet-funding.md. User-only mutation recipes (policy attachment, owner registration): https://github.com/ProjectOpenSea/opensea-skill/blob/main/docs/policy-administration.md.

60-second setup

1. Create a Privy app at dashboard.privy.io; copy App ID + App Secret. Paste both into Pinata's env UI.
2. Deploy. That's the entire required setup.

On first conversation the agent will walk you through:

• Fetching an OpenSea API key and attaching it (free-tier instant key, no signup).
• Creating your Privy server wallet (opensea wallet create) and storing the ID.
• Generating its own additional_signer keypair and storing the private half as PRIVY_AUTH_SIGNING_KEY.
• An off-machine ceremony you do on your own host: generate an owner keypair, then register both your owner public key (as owner_id) and the agent's additional_signer public key on the wallet via https://github.com/ProjectOpenSea/opensea-skill/blob/main/docs/policy-administration.md. Your owner private key never touches the agent.
• Choosing and attaching a per-tx policy (Agent Trading — Conservative is the default template).
• Funding the agent wallet to a hot-wallet float you set per chain.

Each step that needs new secrets ends with a Pinata restart, after which the agent resumes from where you left off. BOOTSTRAP is a resumable state machine — cold restarts pick up at the first incomplete phase.

Example prompts

"Add boredapeyachtclub, pudgypenguins, and azuki to my watchlist on ethereum. Alert me on any floor drop over 5%."

"Follow vitalik.eth and 0xpunks4156 as whales — high-priority alerts."

"Should I buy this azuki at 2.1 ETH? Run the full gate."

"Any upcoming drops this week that fit my taste?"

"What's the best offer on my bored ape #1234? Is it worth flipping?"

"Swap 0.05 ETH into USDC on Base."

"Walk me through tightening the wallet policy to cap buys at 0.1 ETH."

Safety model

Three independent layers, in order of how the bound is actually enforced:

1. Wallet float — the real aggregate ceiling. The agent doesn't hold a key to a treasury; it can only spend what's in the agent wallet. You fund it from your own cold/funding wallet to ≈ a day or week of intended budget and replenish on your cadence. Privy can't enforce daily/weekly cumulative limits, so wallet balance is what stops a runaway spend.
2. Privy per-tx policy (TEE-enforced) — caps each individual transaction. Conditional on owner_id being registered: the agent's env credentials cannot rewrite the policy because mutations require an authorization signature from the owner key (which lives on your host, not the agent's). Bootstrap verifies this is in place before any signing-capable step.
3. Per-turn confirmation for material actions. Any buy, offer acceptance, approval, or transfer above confirmAboveEth (in workspace/TOOLS.md) needs explicit "yes" in the current turn. Snipes can bypass this only when the listing is fully inside your configured envelope — see workspace/SOUL.md → Hierarchy of Ceilings.

Other invariants:

• Env-only credentials. No private keys in the repo or agent workspace. The owner key never touches the agent host.
• Sell-side always confirms. Privy's native-value cap is denominated in the chain's native token (ETH, etc.) — it does not apply to WETH transfers, which is how Seaport offer acceptances pay out. Sells, offer acceptances, and ERC721/1155 approvals therefore always require per-turn confirmation, regardless of price.
• Pre-Buy Gate. Wash trades, thin markets, uneconomic gas, fee surprises, and float exhaustion all block a buy before it's proposed.
• Policy rejections surface verbatim. No workarounds.
• Forbidden operations. The agent will refuse to call any Privy admin endpoint (policy mutation, owner key rotation, auth key changes) — even if you ask. To change the cap or the policy, do it yourself on your machine via https://github.com/ProjectOpenSea/opensea-skill/blob/main/docs/policy-administration.md.

Repository layout

.
├── manifest.json              # Pinata agent manifest — attaches opensea-marketplace from ClawHub
├── LICENSE                    # MIT
├── .openclaw/
│   ├── openclaw.json          # OpenClaw harness config (compaction, concurrency)
│   └── SOUL.md                # short canonical persona — points at workspace/SOUL.md
└── workspace/
    ├── SOUL.md                # guardrails + Conviction Score + Pre-Buy Gate
    ├── AGENTS.md              # workspace conventions + memory schemas
    ├── IDENTITY.md            # blank — filled on first run
    ├── TOOLS.md               # watchlist, whales, budgets — user-tunable
    ├── BOOTSTRAP.md           # first-run walkthrough — agent deletes after completion
    ├── HEARTBEAT.md           # idle-cycle routine
    ├── USER.md                # collector profile — filled on first run
    └── memory/                # created at runtime — floors, actions, taste, scan state

At deploy time Pinata attaches the OpenSea skill under skills/opensea/ (SKILL.md + references/*.md + scripts/*.sh) — not checked into this repo.

Updating the skill

Skill versions are managed on ClawHub, not in this template. To pick up a new version, no repo change is needed — Pinata pulls the latest published version of opensea-marketplace on each deploy. To pin to a specific version, replace clawhub_slug in manifest.json with a cid for that version.