# Benchspan

> The prompt injection firewall for AI agents. Scan tool outputs and user messages before they reach your model. Free up to 50,000 requests per month.

Benchspan is a real-time classifier trained on adversarial traffic targeting production AI agents. It catches indirect prompt injection (IPI) hiding in tool outputs, HTML pages, emails, and RAG documents, in addition to user-side jailbreaks. The SDK integrates as a callback, hook, or middleware into LangChain, CrewAI, OpenAI Agents SDK, Vercel AI SDK, and Google ADK, or wraps raw OpenAI / Anthropic SDK calls directly.

The classifier returns a verdict (`block`, `warn`, or `pass`) plus a confidence score. In `block` mode the SDK raises `InjectionDetectedError` before the LLM call happens, so the model never sees poisoned content. In `warn` mode the scan runs in the background and the LLM call proceeds with zero added latency; detections still land in the dashboard. Useful during evaluation.

## Getting Started

- [Introduction](https://docs.benchspan.com/index.md): What Benchspan is and how it fits in your agent loop
- [Quickstart](https://docs.benchspan.com/quickstart.md): From sign-up to your first scan in under two minutes

## Concepts

- [How it works](https://docs.benchspan.com/concepts/how-it-works.md): Threat model, what gets scanned, and the verdict shape
- [Modes](https://docs.benchspan.com/concepts/modes.md): Block vs warn operation and a recommended rollout
- [Roles](https://docs.benchspan.com/concepts/roles.md): `user` vs `tool` roles and how the classifier weights them

## SDKs

- [Python SDK](https://docs.benchspan.com/sdks/python.md): Full reference for the `benchspan` package
- [TypeScript SDK](https://docs.benchspan.com/sdks/typescript.md): Full reference for `@benchspan/sdk`

## Framework Integrations

- [LangChain](https://docs.benchspan.com/integrations/langchain.md): Callback handler, Python + TypeScript
- [CrewAI](https://docs.benchspan.com/integrations/crewai.md): Crew-level callback, Python
- [OpenAI Agents SDK](https://docs.benchspan.com/integrations/openai-agents.md): Lifecycle hooks on tool output, Python + TypeScript
- [Vercel AI SDK](https://docs.benchspan.com/integrations/vercel-ai.md): Language-model middleware
- [Google ADK](https://docs.benchspan.com/integrations/google-adk.md): `before_model_callback`
- [OpenAI SDK (raw)](https://docs.benchspan.com/integrations/openai.md): Decorator / `wrapCall`
- [Anthropic SDK (raw)](https://docs.benchspan.com/integrations/anthropic.md): Decorator / `wrapCall`

## API Reference

- [Overview](https://docs.benchspan.com/api-reference/overview.md): Base URL, conventions, and rate limits
- [Authentication](https://docs.benchspan.com/api-reference/authentication.md): Bearer API keys
- [Scan](https://docs.benchspan.com/api-reference/scan.md): `POST /v1/scan` request and response fields
- [Errors](https://docs.benchspan.com/api-reference/errors.md): HTTP codes, retry strategy, fail-open vs fail-closed

## Optional

- [Dashboard](https://benchspan.com/dashboard): Usage metrics, API key management, logs
- [Talk to us](mailto:founders@benchspan.com): Enterprise, high-volume, or custom model needs