A trading desk where six of the seven employees are LLMs

What it does

HyperGrok is a Hyperliquid trading desk built for AI agents — seven role profiles, seventeen skills in the portable SKILL.md format, and read-only Python tools. It runs in Grok Bot, Claude Code, Cursor, Codex, Gemini CLI and Grok Build. The desk starts keyless: market data, funding, open interest and book depth come from Hyperliquid’s public /info endpoint with no wallet needed. Execution only happens after you supply a trade-only API wallet and approve an exact ticket by id.

The interesting bit

The separation of duties is the whole point. Six Bots sit in a shared Trading Floor chat and can only read; one Execution Trader holds the keys and sends each approved order exactly once. The approval gate deliberately lives outside the chat — in a Grok Bot Auto-review “Ask first” rule — because, as the README puts it, a Bot that can read an approval could also write one. That is a threat model most agent projects still pretend doesn’t exist.

Key highlights

• Seven roles with a fixed trade lifecycle: idea → evidence → risk sign-off → your approval → one send → reconciliation → review.
• The Risk Manager sizes trades on a stressed stop — assumed adverse fills plus taker fees on both legs — and can refuse you.
• opening_bell.py and desk_doctor.py run on Python’s standard library alone, with --schema/--json for automation and explicit reporting of partial depth.
• research_dataset.py collects verifiable history bundles (captured responses, hashed manifest) and never fills gaps with synthetic data.
• Your limits file can only tighten the desk’s caps, never loosen them; testnet rehearsal comes before mainnet.

Caveats

• The multi-Bot desk experience is Grok Bot-specific; other tools get the skills but not the Trading Floor group chat.
• Stressed-stop sizing is an assumption — the README is candid that actual losses can exceed it.
• The README’s worked ticket is an illustrative example, not a live trade.

Verdict

Worth a look if you trade on Hyperliquid and want agents with real checks instead of a single prompt with your keys. If you wanted a fire-and-forget autonomous bot, this is explicitly the opposite — and that’s the pitch.