Agent
The agent turns a goal in plain language into a rules-based strategy you can inspect, edit and test. It proposes; you decide.
How a request becomes a proposal
- You describe a goal — budget, focus, how hands-off you want to be.
- The agent drafts a strategy. It has exactly one tool,
propose_strategy, and can use any block Studio supports, including the trend filter. - The policy engine checks it with your limits — the same deterministic checks Studio uses. If anything fails, the agent gets the failures back and one chance to revise.
- It's backtested on on-chain price history with costs.
- You decide: save it to Studio, edit it first, or save and paper trade it.
If a request breaks your limits — say, more than your starting budget — the agent explains what it changed and proposes the closest version that fits.
Guardrails
- Can't touch funds. The agent never sees your keys and can't sign or send anything.
- Your limits apply. Proposals are checked against the limits in Settings.
- Logged. Every proposal and its policy result are written to an audit log.
- No promises. It explains risks, labels backtests as hypothetical, and never predicts returns.
Live on-chain execution isn't enabled yet. Today the agent works with saved strategies and paper trading.
Model
The agent runs through OpenRouter. The default model is anthropic/claude-sonnet-5.5. It only sees the list of priced assets, your limits and your conversation — never your wallet.
Tips
- Be specific about budget, recurring amounts and position caps.
- Ask follow-ups in the same conversation to refine a proposal.
- Ask for a defensive rule ("move to cash when QQQ falls below its 20-day average") to get a trend filter.