Rebalance

Demo

About

Privacy preserving non custodial local rebalancer built for AI agents. You are able to talk to OpenAI's Codex or Claude Code or OpenCode (which supports local LLMs for maximum privacy and sovereignty) to set targets for your portfolio with a simple pie chart UI on the side. Robinhood chain is used to hold your assets, which means the RWAs on there are supported in addition to crypto native assets, and Uniswap is used to do the rebalance swap, all in one transaction via 7702. The UI and wallets are all locally on device and non custodial, so there is no need to give any data or custody rights away. It even works with server wallets like Privy and hardware wallets like ledger. The Chat UI on the side makes it very user friendly to interact with your portfolio for complicated requests like rebalancing, and simple things like starting and stopping the rebalancer have old school button options when that's simpler than chatting. The split is done thoughtfully and with intentions. Share functionality is also implemented such that you can copy your settings/target and send to your friend. All they have to do is input into their chat and they can apply it to their own portfolio. Rebalancing can be done via setting the target manually, or setting your risk tolerance via things like Sharpe ratio while talking to the LLM. It's rebalancing will out survive a given closed chat if you don't manually stop it and you can attach a chat to it later. You can also set gas fee limits, drift required before a rebalance, and how often drifts are checked.

How it's made

The project has a local TypeScript Daemon that runs when you do the /rebalance skill and attaches itself to your local codex/claude code/opencode chatbox. It stores the simple wallet in a secure Mac keystore. Privy wallet uses https://agents.privy.io/ wallets, and ledger attaches to the local hardware via the ledger SDK. The UI is also pure HTML/JS/CSS, and ran locally to preserve max user sovereignty. Robinhood Chain is used to hold the assets with its mainnet public RPC used directly. A user needs no API keys. Simple7702Accounts and its ledger integration is used to batch and make the rebalance transaction in one step, so many approval/swaps are not needed. Ledger clear signing is also implemented in the emulator and can later be promoted to production. Hooks and skills are used to both interact with the model and keep certain input deterministic. For example a share text pasted into model chat will pass through to the daemon directly and automatically be applied without needing to waste LLM tokens. The LLM is only notified and directly makes decisions when needed to save tokens, and everything else is done deterministically via daemon, hooks, skills as much as possible.