From local install to the first agent transaction.

Install everything in one command

The setup command installs the service binary, the complete Agent Skill, starts the local service and opens the setup flow. It supports Windows, macOS and Linux on x64 and arm64.

$ npx @bitslabsec/agentshield setup

Install only the Skill

The default installs all supported targets. Use --target when you only want one Agent environment.

$ npx @bitslabsec/agentshield skill install
$ npx @bitslabsec/agentshield skill install --target codex
$ npx @bitslabsec/agentshield skill install --target claude
$ npx @bitslabsec/agentshield skill install --target agents

The bundle contains SKILL.md, references/api.md, scripts/agentshield.mjs and Agent metadata. The installer verifies the required files before reporting success.

Give the Agent two things

The authorization token is intentionally visible to its assigned local Agent. It replaces private-key access with limited, revocable capabilities. Do not send it to a different Agent or a remote service.

Command reference

Make an authorized request

Use the SDK when integrating AgentShield into an application. Keep the authorization token local and use a stable idempotency key for every operation.

import { AgentShieldClient } from "@bitslabsec/agentshield";

const client = new AgentShieldClient({
  baseUrl: "http://127.0.0.1:9400",
  token: process.env.AGENTSHIELD_GRANT,
});

const operation = await client.send({
  chainId: 8453,
  to: "0x…",
  token: "USDC",
  amount: "10000000",
  idempotencyKey: "invoice-42-send-1",
  reason: "Pay invoice 42",
});

Operations available to Agents

Built-in EVM mainnets

Public RPC failover is available by default. Custom HTTPS endpoints can be supplied through AGENTSHIELD_RPCS.

Grant capabilities, not custody

Every Agent authorization is bound to one wallet. Configure only the scopes required by the task, then optionally constrain networks, per-operation USD value, daily operation count, recipients, tokens, contracts, expiry and approval behavior.

What happens when policy needs a human

Requests enter pending_approval when the authorization requires review or when pricing cannot be trusted. Review the decoded operation, amount, recipient, network, contract and policy reason in the AgentShield console before approving or rejecting it.

The key stays outside the Agent boundary

• The account key is split with 2-of-3 Shamir Secret Sharing.
• The complete key is reconstructed only inside the local signer for an authorized signing operation.
• Temporary key material is cleared immediately after signing; the wallet session can remain unlocked until its timeout.
• Agents receive revocable authorization tokens—not the mnemonic, private key, PIN or shards.
• Importing a wallet does not invalidate copies of the original mnemonic or private key.

Common checks