overwing weekly downloads statistics

<p align="center"> <a href="https://overwing.ai"> <picture> <source media="(prefers-color-scheme: dark)" srcset="assets/wordmark-dark.svg"> <img src="assets/wordmark.svg" alt="Overwing" width="220"> </picture> </a> </p> <p align="center"><strong>Guardrails for LLM output, in one line.</strong><br> Vercel AI SDK middleware, OpenAI Agents SDK guardrails, and a typed client. Also: <a href="#atlas-who-is-this-user-agent-no-key-needed">Atlas</a> user-agent lookups with no key, <a href="#preflight-should-the-agent-sign-this-solana-transaction">Preflight</a> for checking a Solana transaction before an agent signs it, and <a href="#tower-let-an-agent-operate-a-legacy-system">Tower</a> for agents operating legacy systems. Every model response gets a <code>pass</code> / <code>fail</code> / <code>review</code> verdict with calibrated confidence before it reaches your user.</p> <p align="center"> <a href="https://www.npmjs.com/package/overwing"><img alt="npm" src="https://img.shields.io/npm/v/overwing?color=0B1220&label=overwing"></a> <a href="https://github.com/frod27/overwing-js/actions"><img alt="CI" src="https://github.com/frod27/overwing-js/actions/workflows/ci.yml/badge.svg"></a> <a href="https://overwing.ai/docs"><img alt="API reference" src="https://img.shields.io/badge/API-reference-0B1220"></a> <a href="https://overwing.ai"><img alt="agents welcome" src="https://overwing.ai/badge.svg"></a> </p>

---

npm install overwing

It works with no key: new Overwing().evaluate(text) runs 10 evaluations a day on inputs up to 2,000 characters, and text sent without a key is not stored. For more, get a free API key at overwing.ai (250 evaluations a day), or let your agent make its own account with Overwing.signup(), which needs no email (see An account for an agent).

The text can be in any language. It was tested on 2026-09-29 in Spanish, Portuguese, French, German, Japanese, Simplified Chinese, Korean, Arabic and Hindi: a small test, not a benchmark. Results come back in English.

Vercel AI SDK middleware

Wrap any model. Works with AI SDK 6 and 7 (generateText, streamText, generateObject, agents, everything).

import { openai } from "@ai-sdk/openai";
import { generateText, wrapLanguageModel } from "ai";
import { overwingGuardrail } from "overwing/ai-sdk";

const model = wrapLanguageModel({
  model: openai("gpt-5"),
  middleware: overwingGuardrail(), // reads OVERWING_API_KEY
});

const { text, providerMetadata } = await generateText({ model, prompt: "Draft a reply to this customer." });
console.log(providerMetadata?.overwing); // { verdict: "pass", confidence: 0.97, id: "eval_…", … }

What happens by default:

Verdict	Default action	Change it with
pass	Response returned, verdict attached to providerMetadata.overwing
review	Returned and annotated so you can route it	onReview: "throw" | "replace"
fail, recommended action block	Throws OverwingGuardrailError (carries the full evaluation)	onFail: "replace" | "annotate"
fail, recommended action redact (e.g. personal data)	Replaces the text with the safe fallback instead of throwing	honorActions: false to treat every fail alike

Streaming is buffered by default: tokens are held until the model finishes and the verdict is in, so a failing response never reaches the screen. Set streaming: "passthrough" to stream immediately and end with an error on fail.

overwingGuardrail({
  ruleSet: "content-safety",        // or your own rule set slug
  onFail: "replace",                // "throw" (default) | "replace" | "annotate"
  onReview: "annotate",             // "annotate" (default) | "throw" | "replace"
  replacement: "I can't share that response.",
  checkInput: true,                 // also score the user's latest message before calling the model
  streaming: "buffer",              // "buffer" (default) | "passthrough"
  metadata: (params) => ({ userId: "u_123" }),
  context: (params) => ({ recipient: "customer", channel: "chat" }),   // facts the rules may reference
  honorActions: true,               // redact-level fails replace rather than throw
  onVerdict: (evaluation, phase) => console.log(phase, evaluation.verdict),
  failOpen: false,                  // true = if Overwing is unreachable, let the response through unscored
});

OpenAI Agents SDK guardrails

import { Agent, run } from "@openai/agents";
import { overwingInputGuardrail, overwingOutputGuardrail } from "overwing/openai-agents";

const agent = new Agent({
  name: "Support",
  instructions: "Help the customer.",
  inputGuardrails: [overwingInputGuardrail()],     // scores the user's message
  outputGuardrails: [overwingOutputGuardrail()],   // scores the agent's final answer
});

try {
  const result = await run(agent, "Reach me at dana@example.com to sort out the refund.");
} catch (err) {
  // InputGuardrailTripwireTriggered / OutputGuardrailTripwireTriggered
  // err.result.output.outputInfo.evaluation → the full Overwing verdict
}

Both accept ruleSet, tripOn: "fail" | "fail-or-review", metadata, onVerdict, and failOpen. Input guardrails run in parallel with the agent by default; pass runInParallel: false to block before the model is called.

Atlas: who is this user agent? (no key needed)

Overwing Atlas is a registry of AI crawlers, fetchers and browser agents. Give it a User-Agent string and it says what the string claims to be and whether that claim can be trusted.

import { Atlas } from "overwing";

const atlas = new Atlas(); // no key: 10 lookups a day. With OVERWING_API_KEY set: 100, or your Atlas plan's limit.…