vouch-guard

Governance guard for Claude Managed Agents. Evaluates every tool call before execution.

ACCEPTED flows. RESTRICTED steers. BLOCKED stops.

Install

npm install vouch-guard

Quick Start

import Anthropic from "@anthropic-ai/sdk";
import { guardSession } from "vouch-guard";

const client = new Anthropic();

// Create agent and session (see Anthropic docs)
const session = await client.beta.sessions.create({ agent: agentId, environment_id: envId });

// Send the task
await client.beta.sessions.events.send(session.id, {
    events: [{ type: "user.message", content: [{ type: "text", text: "Deploy the latest build" }] }],
});

// Guard the session — every tool call evaluated by Vouch
const audit = await guardSession(client, session.id, {
    apiKey: process.env.VOUCH_API_KEY,
    onVerdict: (v) => console.log(`${v.tool}: ${v.verdict} (${v.duration_ms}ms)`),
});

console.log(`Session complete. ${audit.governed} calls governed, ${audit.blocked} blocked.`);

Two Patterns

Pattern 1: Event Monitor (built-in tools)

Guard built-in tools (bash, write, edit, web_fetch). Governance is reactive — the tool may start executing before the verdict arrives. If RESTRICTED or BLOCKED, the guard steers the agent away from repeating the action.

const audit = await guardSession(client, sessionId, {
    apiKey: process.env.VOUCH_API_KEY,
});

Pattern 2: Custom Tool Wrapper (full prevention)

Replace built-in tools with governed custom tools. The tool does NOT execute until Vouch clears it.

import { guardSession, GOVERNED_TOOL_DEFINITIONS } from "vouch-guard";

// Create agent with governed tools instead of built-ins
const agent = await client.beta.agents.create({
    name: "Governed Assistant",
    model: "claude-sonnet-4-6",
    tools: [
        {
            type: "agent_toolset_20260401",
            default_config: { enabled: false },
            configs: [
                { name: "read", enabled: true },
                { name: "glob", enabled: true },
                { name: "grep", enabled: true },
            ],
        },
        ...GOVERNED_TOOL_DEFINITIONS,
    ],
});

// Guard with a tool executor
const audit = await guardSession(client, sessionId, {
    apiKey: process.env.VOUCH_API_KEY,
    executeTool: async (name, input) => {
        // Your execution logic here
        if (name === "governed_bash") {
            return execSync(input.command).toString();
        }
    },
});

API

guardSession(client, sessionId, options)

Guards a managed agent session. Returns an audit summary when the session goes idle.

Options:

| Option | Type | Required | Description | |--------|------|----------|-------------| | apiKey | string | Yes | Vouch API key | | apiUrl | string | No | Vouch API URL (default: vouch.atlaswithiris.com) | | policy | object | No | Custom governance policy | | onVerdict | function | No | Callback on each verdict | | onError | function | No | Callback on Vouch errors | | executeTool | function | No | Handler for custom tool execution (Pattern 2) | | failOpen | boolean | No | Fail open if Vouch unreachable (default: true) |

Returns: Audit summary object.

extractPlan(name, input)

Convert a tool call into a Vouch plan description.

GOVERNED_TOOL_DEFINITIONS

Pre-built custom tool definitions for governed_bash, governed_write, governed_edit, governed_fetch. Spread into your agent's tools array.

VouchClient

Thin API client for POST /api/v1/vouch.

License

MIT