AgentShield

LOCAL-FIRST AGENT PAYMENTS

Agent-powered transactions.
Human-controlled custody.

AgentShield turns an EVM account into a policy-controlled payment sandbox. Agents can sign, send, swap and bridge—without ever receiving your private key or wallet password.

Install in one command View source
  • ✓ Runs on your machine
  • ✓ Scoped, revocable grants
  • ✓ Human approval when needed
QUICK STARTmacOS · Linux · Windows
npx @bitslabsec/agentshield setup
AgentShield Local service connected
ACTIVE GRANTTreasury Agent
LIVE
SENDSWAPBRIDGEAPPROVE
PER OPERATION
$250 USD
NETWORKS
3 allowed
EXPIRES
42m 18s
↗
POLICY CHECKSend 125 USDC on Base
APPROVED
9EVM mainnets
2-of-3Shamir key shards
0Keys shared with agents
100%Local policy enforcement

A BETTER PAYMENT BOUNDARY

Autonomy without unrestricted custody.

Ordinary wallets force a bad choice: give an agent the key, or make every action manual. AgentShield introduces a local control plane between intent and execution.

01 / INTENT

Your agent asks for an outcome

Agent

Send 125 USDC on Base to the approved vendor.

  • Natural-language task
  • No private key access
  • Stable idempotency key
02 / AGENTSHIELD

Local policy decides what happens

✓Scopewallet:send
✓Limit$125 / $250
✓NetworkBase allowed
✓RecipientAllowlisted
03 / RESULT

A signed, traceable operation

Transaction submitted0xa93f…12c8
  • Key reconstructed only for signing
  • Operation tracked to confirmation
  • Complete local audit trail

SECURITY MODEL

Built around the moment of signing.

AgentShield does not pretend a key is never reconstructed. It limits reconstruction to the signing boundary, keeps the agent outside that boundary, and clears the complete key from memory immediately after use.

01
2 / 3

Encrypted key shards

The account key is split with 2-of-3 Shamir Secret Sharing. Local and recovery material are protected with Argon2id and AES-256-GCM.

02
RAM

Ephemeral reconstruction

The complete account key exists only inside the signer for the operation. It is cleared after signing while the authorized wallet session remains usable.

03
POLICY

Capabilities, not secrets

Agents receive a revocable token constrained by scopes, USD limits, networks, recipients, assets, contracts, expiry and approval requirements.

LOCAL-FIRST

Wallet passwords, shards and signing material stay on the machine running AgentShield. Agents communicate through a narrow HTTP API and never receive wallet credentials.

ONE BOUNDARY, COMPLETE WORKFLOWS

Everything an agent needs to pay.

Each capability is separately authorized. Start read-only, add exactly what the task needs, and revoke access at any time.

SIGN

Messages & payloads

Sign human-readable messages or complete structured EVM transactions with arbitrary calldata after policy review.

SEND

Native & ERC-20

Transfer ETH, stablecoins and exact-address tokens with recipient, network and USD controls.

SWAP

Same-chain exchange

Discover assets, request LI.FI quotes and execute policy-checked swaps without handing custody to the agent.

BRIDGE

Cross-chain movement

Route funds across supported EVM networks and track every operation through final confirmation.

APPROVE

Delegated review

Authorize a trusted agent to inspect and approve same-wallet requests without escaping the parent policy.

DELEGATE

Child grants

Create narrower, shorter-lived grants. Child permissions can never exceed the parent authorization.

EVM-FOCUSED

Nine mainnets, one policy surface.

Public RPC failover works out of the box. Exact custom EVM endpoints can be added when your infrastructure requires them.

Ethereum 1Optimism 10BNB Chain 56Polygon 137zkSync Era 324Robinhood Chain 4663Base 8453Arbitrum 42161Avalanche 43114

FROM WALLET TO AGENT IN MINUTES

One local setup. A clear handoff.

  1. 01

    Create or import an account

    Generate a fresh EVM account or import a mnemonic/private key into the local conversion flow. Each wallet keeps its own password and grants.

  2. 02

    Define the operating envelope

    Choose low, medium or high presets—or set exact scopes, limits, allowlists, expiry and approval behavior.

  3. 03

    Give the agent one Skill and one token

    The Skill explains the complete API. The scoped token supplies only the capabilities you selected. The agent can verify its connection itself.

SHIP THE BOUNDARY, NOT THE KEY

Ready in one command.

The npm package installs the service and complete Agent Skill for Codex, Claude Code and compatible Agents on Windows, macOS and Linux.

terminal
$ npx @bitslabsec/agentshield setup
Documentation →npm package ↗GitHub repository ↗

QUESTIONS

What teams ask first.

Does the agent ever see my private key?

No. The agent receives a scoped authorization token. The signer briefly reconstructs the account key in local memory for signing and clears it immediately afterward.

Can an agent transact without human approval?

Yes, when its grant explicitly allows direct execution and the request passes every policy check. Unknown token prices and out-of-policy requests remain blocked for human review.

Is AgentShield a cloud wallet?

No. It is a local-first sandbox and signing service. Wallet secrets and policy enforcement remain on the machine where AgentShield runs.

Can I manage more than one account?

Yes. AgentShield supports up to ten wallets, each with independent passwords, key shards, grants and activity history.