Skip to main content

​
Overview

@sigilcore/agent-hooks is the client-side enforcement layer for Sigil. It intercepts an agent’s intended tool call before it executes, submits it to the Sigil Sign /v1/authorize endpoint, and blocks or holds the action based on the policy decision. Without agent-hooks, Sigil Sign governs EVM transactions only. With agent-hooks, Sigil governs any agent action on any framework — bash commands, HTTP requests, file writes, wallet signing, and email sends. The TypeScript package is the JavaScript integration surface. Rust hosts use the companion agent-hooks-rs crates, which share the same /v1/authorize wire fixtures and add a native IronClaw hook adapter.

​
Installation

npm install @sigilcore/agent-hooks

​
How It Works

Every tool call an agent attempts is intercepted before execution: 
Agent attempts tool call
        ↓
@sigilcore/agent-hooks
        ↓
POST /v1/authorize → Sigil Sign
        ↓
Policy evaluated against warranty.md
        ↓
APPROVED → tool executes
DENIED   → typed rejection returned to agent
PENDING  → action held for human approval

​
Supported Frameworks

FrameworkIDPackageAdapter
Generic TypeScript hostagent-hooks@sigilcore/agent-hookscheckIntent
Claude Code / Anthropic SDKanthropic-sdk@sigilcore/agent-hookscheckAnthropicToolUse
ELIZAeliza@sigilcore/agent-hookscheckElizaAction
LangChainlangchain@sigilcore/agent-hookswrapLangChainTool
OpenClawopenclaw@sigilcore/agent-hookscreateOpenclawSigilHandler
NVIDIA NemoClawnemoclaw@sigilcore/agent-hookscreateOpenclawSigilHandler
IronClawironclawsigil-agent-hooks-ironclawnative Rust Hook
USD1 AgentPay (WLFI)agentpay@sigilcore/agent-hookshost-level checkIntent wrapper
Any frameworkcustomTypeScript or Rustgeneric client call
See the Framework Registry for the full list and custom framework usage.

​
Governed Actions

ActionDescription
bashShell command execution
web_fetchOutbound HTTP requests
file_writeFilesystem writes
wallet_signEVM wallet signing
email.sendOutbound email
wallet.transferEVM token transfers
contract.callEVM contract calls

​
Prerequisites

You need a Sigil API key and a signed warranty.md policy file deployed to Sigil Sign.

​
Fail Modes

When Sigil Sign is unreachable, agent-hooks can either fail open or fail closed. Unreachability includes network errors, DNS failures, refused connections, request timeouts, 5xx responses, and non-JSON response bodies.

​
TypeScript: @sigilcore/agent-hooks

The TypeScript package defaults to failMode: 'open' for backward compatibility with v0.1.0. 
import { checkIntent } from '@sigilcore/agent-hooks';

const result = await checkIntent(intent, {
  apiKey: process.env.SIGIL_API_KEY!,
  agentId: 'production-agent',
  failMode: 'closed',
});
ModeUnreachable resultUse when
failMode: 'open'APPROVED with failOpen: trueLocal development and non-financial workflows
failMode: 'closed'DENIED with SIGIL_UNREACHABLEProduction, externally-visible actions, and wallet or on-chain actions
In open mode, fallback approvals carry failOpen: true so hosts can distinguish an outage fallback from a real policy approval. In closed mode, buildRejectionContext tells the agent to pause and retry after connectivity is restored; it does not frame the event as a policy violation.

​
Rust: agent-hooks-rs

The Rust crates default to FailMode::Closed because they have no legacy fail-open behavior to preserve. They expose FailMode::Open for development or low-risk workflows.

Rust and IronClaw

Use sigil-agent-hooks-core directly from Rust or sigil-agent-hooks-ironclaw as a native IronClaw Hook.

​
Source