@rends/agent-sdk

Runtime interception SDK for AI agent governance. Wraps agent tool calls with synchronous policy enforcement via the AgentCompliant/Rends platform.

Install

# TypeScript / Node.js
npm install @rends/agent-sdk

# Python
pip install rends-agent-sdk

Quick Start (TypeScript)

import { RendsClient } from '@rends/agent-sdk';

const client = new RendsClient({
  apiKey: 'ac_live_...',
  orgId: 'your-org-uuid',
  agentId: 'your-agent-uuid',
  mode: 'enforce',  // 'enforce' | 'monitor' | 'dry-run'
});

// Wrap any tool call with governance
const result = await client.govern(
  {
    actionType: 'database_query',
    actionName: 'select_users',
    resourceType: 'customer_records',
    inputSummary: 'SELECT * FROM users WHERE region = us-east-1',
    metadata: { dataClassification: 'pii' },
  },
  async (params) => db.query('SELECT * FROM users WHERE region = $1', ['us-east-1'])
);

// result.allowed = true/false
// result.status = 'pass' | 'warn' | 'block'
// result.result = query output (if allowed)
// result.check.violations = [] (if blocked, why)

// Flush telemetry before exit
await client.drain();

Quick Start (Python)

from rends_sdk import RendsClient, RendsClientConfig, ActionContext

async with RendsClient(RendsClientConfig(
    api_key="ac_live_...",
    org_id="your-org-uuid",
    agent_id="your-agent-uuid",
)) as client:
    result = await client.govern(
        ActionContext(
            action_type="database_query",
            action_name="select_users",
            resource_type="customer_records",
            input_summary="SELECT * FROM users WHERE region = us-east-1",
        ),
        tool_fn=lambda params: db.query("SELECT * FROM users"),
    )

LangChain Integration

import { RendsClient } from '@rends/agent-sdk';
import { governTool, governTools } from '@rends/agent-sdk/adapters/langchain';

const client = new RendsClient({ apiKey: '...', orgId: '...', agentId: '...' });

// Wrap a single tool
const governedSearch = governTool(client, searchTool, 'web_search');

// Wrap all tools at once
const governedTools = governTools(client, [search, calculator, browser], 'tool');

// Use with any LangChain agent
const agent = createReactAgent({ llm, tools: governedTools });

CrewAI Integration (Python)

from rends_sdk.adapters import govern_crewai_tool

search = SerperDevTool()
governed = govern_crewai_tool(client, search, resource_type="web_search")
agent = Agent(tools=[governed], ...)

AutoGen Integration (Python)

from rends_sdk.adapters import govern_autogen_tool

def search_web(query: str) -> str:
    return tavily.search(query)

governed = govern_autogen_tool(client, search_web, "search_web", resource_type="web_search")
agent.register_function(function_map={"search_web": governed})

Enforcement Modes

| Mode | Behavior | |------|----------| | enforce | BLOCK → throws GovernanceBlockError, PASS/WARN → executes | | monitor | All decisions logged but never block execution | | dry-run | Policy checked, nothing executed, nothing logged |

Pre-flight Check

// Check policy without executing
const check = await client.checkPolicy({
  actionType: 'data_export',
  actionName: 'export_pii',
  resourceType: 'user_data',
  inputSummary: 'Export all PII to CSV',
});
// check.allowed, check.status, check.violations

Pipeline Stages

The SDK runs a 5-stage pipeline on every govern() call:

1. beforeAction — validates action shape, computes input hash
2. policyCheck — calls POST /v1/compliance/check-action synchronously
3. decisionReturn — parses { allowed, status, violations, warnings }
4. executionGate — executes tool if allowed, throws if blocked (in enforce mode)
5. telemetryEmit — logs to POST /v1/telemetry/ingest (hash-chained audit trail)

API Reference

RendsClient(config)

| Config | Type | Default | Description | |--------|------|---------|-------------| | apiKey | string | required | API key (ac_live_...) | | orgId | string | required | Organization UUID | | agentId | string | required | Agent UUID | | endpoint | string | https://agentcompliant.ai/api | API base URL | | mode | string | enforce | Enforcement mode | | timeoutMs | number | 10000 | Policy check timeout | | retries | number | 2 | Retry count for transient failures | | failOpen | boolean | false | Allow action if policy check fails |

ActionContext

| Field | Type | Max Length | Description | |-------|------|-----------|-------------| | actionType | string | 200 | Category: tool_call, api_request, data_access | | actionName | string | 500 | Specific action: database_query, send_email | | resourceType | string | 200 | Target resource: customer_records, email_server | | inputSummary | string | 8000 | Free-text description of the input | | metadata | object | — | Additional context (not sent to policy check) |