Philosophy
An agent that works for you, inside hard limits it cannot argue with.
Guarded workflow
- 01
Read
Tier 1 tools read markets, policy and token data. Nothing at this step can move funds.
- 02
Prepare and Confirm
In the app the assistant proposes an order, or executes it if you enabled the optional Agent Wallet. Over MCP, Tier 2 tools simulate and prepare it. Either way Zypher Check runs against your limits first.
- 03
Execute and Audit
In the app you approve the proposal in your own wallet, or the Agent Wallet executes it within your caps. Over MCP, execute_intent takes only the confirmationId and checks policy again. Every execution is journaled with a rationale.
Broad market reading. Deep guarded execution. Held together by one policy.
Hard Limits
Limits live in policy.json, outside the model. They are re-read on every call and the agent cannot negotiate them.
One-Time Confirmation
App proposals expire after 10 minutes. Ask mode is the default and Auto mode is opt-in, with per-trade and daily caps, a leverage cap, allowed coins and an instant Pause. MCP confirmationIds expire after 120 seconds and are bound to a hash of their parameters.
Testnet First
The app runs on Elysium and Hyperliquid testnets only, testnet tokens have no value, and mainnet is not enabled. Over MCP, execution is dry-run by default and mainnet needs a policy flag and an environment acknowledgement.
Untrusted Text
Tweets and token names are wrapped as untrusted external text, so outside content cannot become an instruction.
Zypher Agent in Brief
A short overview of the tools, the policy model and the roadmap is available in the app.