Sherwood is in private beta. Vault creation is open to waitlisted creators today, and general access is coming soon. Join the waitlist.
Why most AI hedge fund projects stop at a laptop
Open-source “ai-hedge-fund” repos on GitHub show the research side well: analyst agents, a bull and bear debate, a risk manager and a trader. They run against a broker account or a backtest. What they leave out is the fund itself:- Outside capital. Taking other people’s money needs shares, deposits and redemptions.
- A public record. Depositors want to see what the agent did, not a screenshot of a backtest.
- Visibility before money moves. Depositors want to read a trade before it settles.
The roles inside an AI trading agent
Keep the agent’s job split into clear steps, whatever model you use:- Research. Gather prices, news and market data. Sherwood’s data feeds list Robinhood Chain Stock Tokens, prices and pool liquidity without a login.
- Thesis. Write down why the trade should work, the kill criteria and the planned settle date.
- Sizing. Fit the trade to the caps set on the vault and the risk you will accept.
- Proposal. Turn the plan into a proposal depositors can read.
- Review. Watch the vote, execute, settle, then compare the result to the thesis.
Step 1: Set up the agent
Install the CLI (Node.js 22 or later) and the Sherwood skill pack so the agent knows the commands:https://mcp.sherwood.sh/mcp. It reads vault state and prepares unsigned transactions for your own wallet to sign. To run the agent on Grok, follow Set up Grok to run your fund.
Step 2: Create the fund
Follow Getting started to set up a wallet, mint the agent’s ERC-8004 identity, stake, create the vault and approve your first depositor. Each vault is a standard ERC-4626 vault. Depositors get shares, and deposits and withdrawals follow the vault’s rules.Step 3: Write the mandate
Before the first trade, write down what the fund does: which assets, how much of the vault a single proposal may use, how long positions last and the drawdown you will accept. Set the caps on the vault to match, and publish a short version so depositors know what they signed up for.Step 4: Run the proposal loop
Every trade is a proposal. The agent plans it, sizes it, proposes it, and depositors can veto it during the vote. After execution it runs until it settles. How to run an AI-managed fund walks through every step, including timing around the hours when stock price feeds pause.Step 5: Build the track record
Each proposal, vote and settlement stays on Robinhood Chain, so the fund’s history builds itself. Read it through the SDK, HTTP API or CLI, and point depositors to it instead of a pitch deck.Risks to take seriously
- Model error. Agents can misread data or overtrade. Keep a human reading every proposal at first.
- Bad or stale data. Check feeds before you propose. See Oracle staleness.
- Smart-contract risk. Any contract can have bugs. Community templates are marked as such in the docs.
- Market risk. Tokenized stocks and crypto can fall fast. Size for the drawdown you wrote down.