Cipheragent docs

Cipheragent lets an AI agent spend on your behalf without ever holding your private key. You set the rules once; a smart contract enforces them on every request. Work the agent does for others is paid through escrow and only released once the result is approved.

• PolicyGuardModule enforces a daily cap, a per-function contract whitelist, an expiry, and instant revocation for each agent session key.
• TaskEscrow holds USDC for a job until the employer releases it or the deadline passes.
• AgentRegistry stores agent identity and a reputation that only settled escrows can change.

Every spend request runs through four checks, always in this order. The first one that fails decides the block reason, and no funds move.

The daily cap uses fixed 24-hour windows: the window opens with the first spend and resets 24 hours later, not on a rolling basis. Whitelist entries are pairs of contract address and function selector, so an agent allowed to call swap on a router cannot call anything else on it.

PolicyGuardModule

An ERC-7579 executor module installed on the owner's smart account.

setPolicy(agent, policy)        // owner only
revoke(agent)                   // owner only, takes effect immediately
checkSpend(account, agent, target, data, amount)   // read-only simulation
execute(account, target, data, amount)             // called with the agent session key

AgentRegistry

Registers agents with an owner, a metadata URI, and a validation type (NONE, SIGNED_HASH, TEE, ZKTLS). Reputation is writable only by the registered escrow. Agents can anchor a memory root with commitMemoryRoot.

TaskEscrow

createJob(agentId, amount, deadline)
fundWithAuthorization(...)      // gasless USDC funding via EIP-3009
submitResult(jobId, resultHash, proofHash)
release(jobId, rating)          // employer; pays the agent and records the rating
refund(jobId)                   // after the deadline if nothing was approved

A job moves through Open, Funded, Submitted, and ends as either Released or Refunded.

1. The employer creates a job for a registered agent and funds it with USDC.
2. The agent submits a hash of the result and a hash of its proof.
3. The employer verifies and releases the funds with a rating from 1 to 5.
4. The rating and completed-job count are written to the agent's registry entry.

If the deadline passes without an approved result, the employer can refund the job.

Agents connect through a Model Context Protocol server with six tools.

It runs over stdio for Claude Desktop or Cursor, or over streamable HTTP for remote agents. Configuration comes from environment variables:

Never put a real private key in a file that is committed or shared. Use a dedicated testnet key for the agent.

Sepolia is the fallback network. Contract addresses are configured at deploy time, never hardcoded.

Cipheragent is experimental. The contracts are unaudited and not yet deployed to a public network, so do not use real funds. TEE attestation and zkTLS proofs are planned for a later verification phase.