Crypto Hunter AI

A multi-agent opportunity intelligence organization for crypto — the flagship org of the Veritas Deterministic Scaffold thesis.

Scout agents monitor known sources for legitimate earning opportunities (airdrops, quests, testnets, Learn & Earn). Verification agents investigate. A deterministic scaffold gate — the only trust boundary — decides what gets persisted and shown to the user. Free agents at the edges, one incorruptible gate at the center.

Constitution

1. Never holds keys. Never signs or executes transactions.
2. The trust boundary is deterministic — nothing reaches the user except through the fail-closed gate.
3. Every trusted claim carries provenance (source URL, timestamp, confirmations).
4. Presents intelligence, not financial advice.

Architecture (Phase 1)

• Proposal zone (untrusted): Opus 4.8 scouts with server-side web search/fetch scoped to each beat's allowed domains file candidate OpportunitySpecs. The Source Verification Agent hunts independent confirmations; the Scam Detection Agent hunts negative evidence (scam-report URLs). Filing new evidence resets a record to candidate — verdicts are never stale.
• Scaffold zone (deterministic): engine/gate.py — scam-list, official-domain allowlist, source validity (RDAP age on registrable domains), multi-source confirmation, explorer contract check, credible-scam-report check. Fail closed, monotonic: unvalidatable evidence is excluded, never decisive — except scam taint, which is a hard fail anywhere.
• Scoring rubric: cost/time/safety computed deterministically from the spec and the gate's own evidence; the Ranking Agent proposes only reward-potential (0-40), clamped in code. Total /90.
• On-chain sense: the gate's contract_verified check runs against block explorers — Etherscan v2 when ETHERSCAN_API_KEY is set, otherwise the chain's keyless public Blockscout instance (Ethereum/Base/Optimism/Arbitrum/Gnosis). Unknown chain → fails closed.
• The debate: Advocate / Skeptic / Strategist file stances (endorse/caution/oppose) persisted as transcript — never a verdict. The Skeptic's only teeth are scam reports the deterministic gate counts.
• DataHub: SQLite store of all specs + provenance + outcomes + debate transcript + the activity log (every agent action) + usage log.
• Digest & Mission: deterministic ranked content; the LLM only voices it.
• The dashboard is a living command center. "The Army" tab shows all 15 agents as cards — role, latest action, count today, or "standing by" — plus a live hunt feed of the org working (scouts spotting leads, the Skeptic raising flags, the Gate rendering verdicts). Activity logging is purely observational; it can never change a verdict.

Providers

The product target is the Anthropic Claude API (claude-opus-4-8, server-side web search/fetch, adaptive thinking). A second provider, OpenAI, exists purely as build/dev mode so development can continue when the Anthropic account is unfunded: LLM_PROVIDER=openai in .env runs the same agents on gpt-5-mini via the Responses API + hosted web_search. One contract (engine/agents/runtime.py), two implementations; the gate and every trust decision are provider-independent.

Run

source .venv/bin/activate
# .env at repo root: ANTHROPIC_API_KEY=... (product) / OPENAI_API_KEY=... + LLM_PROVIDER=openai (build)
python cli.py day              # sweep -> verify -> hunt-scams -> gate -> rank -> digest -> mission
python cli.py mission          # tonight's Daily Mission from your profile (config/profile.json)
python cli.py serve            # dashboard at http://127.0.0.1:8010
python cli.py debate           # the recommendation debate (stances → deterministic settle)
python cli.py record <id> --acted --paid --payout 120   # outcome tracking
python cli.py status           # counts + token/cost totals
pytest                         # offline tests incl. the seeded-scam regression

Runs itself daily

A macOS LaunchAgent (~/Library/LaunchAgents/com.moresalamander.cryptohunter.plist) runs scripts/daily_run.sh — the full day-cycle — every morning at 07:00 local, so the dashboard's mission and digest are simply there. Logs land in logs/daily-YYYY-MM-DD.log.

launchctl unload ~/Library/LaunchAgents/com.moresalamander.cryptohunter.plist   # stop it
launchctl kickstart -k gui/$(id -u)/com.moresalamander.cryptohunter             # run now

Built by MoreSalamander — "built by the engineer learning how to build it."