Safety & Disclaimer
The honest boundary of the project: request access, equities-only, not investment advice, no track record, an on-chain module that is on mainnet but unaudited with no timelock yet, and what is explicitly out of scope.
Who is behind Sphynx
The project is published under the name Sphynx. No individual is currently named publicly as its maintainer. That is worth saying plainly rather than leaving you to notice it: for a project whose subject is automated trading, an anonymous team is a legitimate reason for caution, and calling the team "Sphynx" does not answer the question — it restates it. If it matters to you, treat it as unresolved. What we can say is what the anonymity is *not* load-bearing for. Nothing here asks you to trust the people: the code is public and reproducible, the guardrails are enforced by contracts you can read and call rather than by a promise, the vault holds no third-party money, and access is gated precisely so that no one is depositing on the strength of a reputation nobody has established. Verified identity and a legal counterparty will be provided to an auditing firm under engagement. If you are evaluating whether to trust this project, the honest ranking of what should worry you is: no third-party audit first, no track record second, anonymity third. The first two are the ones we would fix first, and are being worked on in that order.
Relationship to Robinhood and Anthropic
Sphynx is an independent project with no relationship to Robinhood Markets, Inc. or to Anthropic beyond being a user of their public interfaces. It is not affiliated with, endorsed by, sponsored by, partnered with, or reviewed by either company. Both names appear throughout this documentation because the project is built on top of them, and describing what it does without naming them would be less honest, not more. Specifically: the desk connects to a Robinhood Agentic account through Robinhood's own MCP server; the vault is deployed to Robinhood Chain, a public chain anyone can deploy to; and the agent itself is Anthropic's Claude, running in Claude Code on your machine. Those are dependencies. None of them is a review, an approval, or a safety guarantee — no one at either company has looked at this code. The practical consequence, stated plainly because it is the thing worth taking away: any confidence you have in Robinhood or in Anthropic should not transfer to Sphynx. The vault is unaudited and holds nothing; the guardrails are enforced by this project's contracts, not by Robinhood's. "Robinhood", "Robinhood Chain", "Claude" and "Anthropic" are the marks of their respective owners and are used here only to identify the platforms this project runs on.
No performance claims
Every number, snapshot, example log, and backtest in this project is illustrative/demo unless you feed it real data. Nothing here represents historical performance. The backtester checks whether rule logic is internally consistent, not whether a strategy is profitable, and feeding it real bars does not create a track record. On-chain figures are either a real read or honestly empty, and today they read zero: no TVL, no depositors, no trades.
Scope
Sphynx is deliberately narrow: equities-only, long-only, USD, human-in-the-loop, Claude-Code-native. Options, futures, and crypto are not supported by the current access surface, and the desk is configured to refuse them (option order tools are deny-ed).
Your responsibilities
- Fund the Agentic account with a small budget you can afford to lose, that is your maximum downside.
- Define your risk caps in `strategies/` before trading; the Risk Manager VETOes if they are unset.
- Monitor it yourself. You own the guardrails and the limits; the agent will not change them.
- Keep the repo private, and never commit secrets or the OAuth token. See Configuration.
- Verify anything the desk surfaces before acting on it. It is a second opinion, not a decision-maker.
Open verification items
Some assumptions still need confirmation against official sources before you rely on them:
- Instrument scope
- Robinhood Agentic Trading is currently equities-only, crypto/options/futures support is not available. The desk is intentionally limited to equities long-only; confirm current scope with Robinhood.
- MCP tool names & behavior
- Verify the real tool names via
/mcporclaude mcp get robinhood-trading, then tighten `.claude/settings.json` to match. Names can change while access is gated. - Legal
- Any feature beyond equities could fall under securities regulation and must pass legal review before it is even considered. No legal review has been completed, and deploying a contract does not resolve regulatory exposure, including the US-person question.
- Third-party audit (on-chain module)
- Not done. Unit tests and a mainnet-fork test are not an audit. Nothing on chain should be relied on until an independent audit exists.
- Explorer verification
- The mainnet contracts are not yet verified on the block explorer, so you cannot yet read the deployed source there. Verify addresses against
onchain/deployments/latest.jsonand by direct call.
Out of scope: crypto / token material
The on-chain module: on mainnet, and unaudited
The on-chain vault + guardrails-as-code module in onchain/ is deployed to Robinhood Chain mainnet, chainId 4663, as of 2026-09-20, against real USDG and the real Uniswap V3 stock-token pools. Every caveat below is current, and the deposit cap stays at 10,000 USDG until the first two are resolved.
- No third-party audit. 29 unit tests and a mainnet-fork round trip are not an audit. This is the single largest caveat on the module.
- No timelock. An owner change (oracle, adapter, guardrails, caps, deposit cap, session grant) takes effect in one transaction. Ownership is a single EOA at launch, which also holds the trading session. A Safe and a timelock are the first planned upgrades; see Risks.
- No track record, no returns, no performance. NAV is 0, there are no depositors, and no trade has been made. Any figure shown anywhere is a real on-chain read, honestly empty, or explicitly labelled sample.
- Deposits are capped at 10,000 USDG, an owner-changeable setting, not structural. Contract source is not yet verified on the block explorer (its API is behind a Cloudflare challenge); verify bytecode against
forge buildoutput. - Prices are a pool TWAP, not an oracle network. There is no Chainlink feed for these tokens on this chain. A 5-minute Uniswap V3 average with a 3% deviation bound is the mark; see Oracle & Execution.
- Legal and securities review is still pending, including the US-person question. The Robinhood Stock Tokens the vault trades are price-tracking instruments, not shares, and are not for US persons. Deploying a contract does not resolve regulatory exposure.
- It is not connected to your Robinhood account and it is not part of the desk's trading path.
- The equities desk is unchanged. It still requires your explicit in-session approval for every order. The mainnet deploy does not make the desk autonomous.
So the desk's scope stays exactly what it is today: equities-only, long-only, human-in-the-loop, Claude-Code-native. Until legal review, a third-party audit, a technical verification of any third-party network, and official confirmation of non-equities availability are all complete, the on-chain module stays a direction, deployed but unaudited code, not a product you can rely on.
Not affiliated
Sphynx is an independent, educational reference architecture built on Claude Code and Robinhood Agentic access. It is not endorsed by or affiliated with Robinhood or Anthropic. Licensed under MIT, see the LICENSE file.
Related reading: Guardrails, Strategies & Risk, and the FAQ.