elytro-agent (CLI shim)

An agent-facing CLI for the elytro-core agent-native smart account. The human delegates once; the agent then operates autonomously within on-chain realized-value caps, and escalates when an action falls outside its envelope. Deterministic JSON output, for agents.

Install

npm i -g @elytro/agent-cli      # provides the `elytro-agent` command

The Cleave testnet (chain 73571, EntryPoint v0.8, free faucet, no real money) is baked in as the default RPC + factory, so a fresh install works with no config. From source: cd cli && npm install && node build.mjs.

Config (env or flags)

| Env | Meaning | |---|---| | ELYTRO_RPC | RPC URL (default: Cleave testnet) | | ELYTRO_FACTORY | AgentAccountFactory address (default: Cleave testnet) | | ELYTRO_ENTRYPOINT | EntryPoint (default 0x4337…F108, v0.8) | | ELYTRO_OWNER_KEY | owner (root) key (human only), used by create/grant. The agent must never hold this. | | ELYTRO_AGENT_KEY | agent session key: signs and self-submits UserOps for send. Generate with keygen (stored at ~/.elytro-agent/agent.key). | | ELYTRO_CHAIN_ID | chain id (default 73571) |

The agent self-submits its own UserOp (acts as its own bundler), so send never needs the owner key, only the agent key plus a little ETH for base gas (refunded to the agent). Every flag (--rpc, --factory, --owner-key, --agent-key, …) overrides its env var.

The agent contract (output)

{ "success": true, "result": { ... } }
{ "success": false, "error": { "code": -32010, "message": "...", "decision": "escalate", "suggestion": "..." } }

decision: "allow" → the agent may act autonomously. decision: "escalate" (-32010) → out of the delegated envelope; get human approval or adjust the grant. -32012 → the action would fail to execute (funding / token). send runs the same simulation as check/simulate and refuses to submit (no gas spent) anything out of envelope.

Claim a transfer succeeded only when result.executed === true and result.userOpSuccess === true. Never infer success from result.status (the bundle tx status is "success" even when the EntryPoint caught an inner revert and the contract refused the operation).

Commands

# (agent) once: generate + store the agent session key, print the agent address
elytro-agent keygen
elytro-agent whoami                       # reprint the agent address anytime

# counterfactual account address
elytro-agent address --owner 0xOwner --salt myacct

# (human/owner) deploy the account
elytro-agent create --salt myacct

# (human/owner) delegate a scoped cap to an agent
elytro-agent grant --account 0xAcct --agent 0xAgent --token 0xUSDC \
  --per-tx 100000000 --total 300000000 --expires-in 2592000

# (agent) dry-run end to end: what would move, would it revert, remaining budget (no broadcast)
elytro-agent simulate --account 0xAcct --token 0xUSDC --to 0xBob --amount 50000000

# (agent) may I do this autonomously? allow | escalate
elytro-agent check --account 0xAcct --agent 0xAgent --token 0xUSDC --amount 50000000 --to 0xBob

# (agent) act: capped transfer via a UserOp through the EntryPoint (add --dry-run to preview)
elytro-agent send --account 0xAcct --token 0xUSDC --to 0xBob --amount 50000000

# read owner, cap, balance
elytro-agent status --account 0xAcct --agent 0xAgent --token 0xUSDC

Verified live (Cleave testnet, chain 73571, real EntryPoint v0.8)

End-to-end through this CLI against real mainnet USDC (0xA0b8…eB48):

• create → account 0xE951eBac98C103e581707198AD9E3c2682A8A41d
• grant → agent cap 100/tx, 300 total
• check 50 → allow; check 150 → escalate (exceeds per-tx cap)
• send 50 → executed (tx 0x6fabae79…), bob +50 USDC, spentTotal 50
• send 150 → refused by the agent's own preflight (-32010), nothing submitted

The realized-value cap is enforced on-chain regardless; the CLI's preflight just saves a doomed transaction.