SDK

Wrap your client. Get tracing, scanning, and a cert.

One line of instrumentation per agent. Works with OpenAI, Anthropic, LangChain, MCP, and any OTel-compatible client.

TypeScript

@vouch/sdk

refund-agent.ts
// @vouch/sdk (TypeScript)
import { instrumentOpenAI, scan, guard } from "@vouch/sdk";
import OpenAI from "openai";

const openai = instrumentOpenAI(new OpenAI(), {
  agentId: "refund-agent",
  baseUrl: process.env.VOUCH_BASE_URL,
});

// 1) auto-instrument every chat call
const reply = await openai.chat.completions.create({
  model: "gpt-4o-mini",
  messages: [{ role: "user", content: input }],
});

// 2) one-off scan
const verdict = await scan({ prompt: input });
if (verdict.action === "block") throw new Error(verdict.reason);

// 3) guard wrapper (throws VouchBlockedError on policy block)
const result = await guard(input, async () => {
  return await callOurAgent(input);
});
Python

vouch_sdk

refund_agent.py
# vouch_sdk (Python)
from vouch_sdk import instrument_openai, scan, guard

openai = instrument_openai(
    openai_client,
    agent_id="refund-agent",
    base_url=os.environ["VOUCH_BASE_URL"],
)

verdict = scan(prompt=user_input)
if verdict.action == "block":
    raise PolicyBlocked(verdict.reason)

result = guard(user_input, lambda: call_our_agent(user_input))
Coverage

What gets instrumented.

  • Trace ingestion — model name, prompt + system, retrieved context, tool calls, tool results, latency, cost, agent fingerprint.
  • Firewall scan — pre-tool input scan + post-generation output scan. Blocks land as findings.
  • Policy gates — guard() wraps any function with the agent's policy graph. Throws VouchBlockedError on block.
  • Cert reporting — every call ticks the score components (firewall block-rate, intent failure-rate). Cert auto-mints every 6h.

Install in 60 seconds.

npm i @vouch/sdkAPI reference