OpenAI Guardrails (JS)
https://github.com/openai/openai-guardrails-js- Category
- AI Tools
- Rank
- No. 1003Tools index
- Pricing
- Open Source
- Type
- TOOL
- Builder
- openai
- GitHub
- 94 stars
- Latest release
- v0.2.1
- Date
About
TypeScript and JavaScript SDK for OpenAI Guardrails — runtime safety controls for LLM applications.
What it does
It wraps OpenAI and Azure OpenAI clients so configured checks can inspect requests and responses at preflight, input, and output stages. Built-in checks cover moderation, URLs, personal data, hallucinations, jailbreaks, topic boundaries, and custom prompts. A separate evaluation workflow measures guardrail behavior against labeled JSONL datasets.
Why it's ranked here
This is a strong fit for teams already using OpenAI clients or the Agents SDK. Client-compatible integration keeps adoption focused, while staged checks, streaming support, configurable tripwires, and dataset evaluation address both deployment and testing. The preview label and default handling of execution failures warrant care in sensitive systems.
What's good
Guardrails run concurrently within a bundle, and configuration passes through schema validation before execution. Responses retain results grouped by pipeline stage, including triggered checks and aggregated token usage. The package also supports conversation-aware checks, periodic validation during streaming, Azure OpenAI, custom registered checks, and evaluation metrics including precision, recall, and F1.
Tradeoffs
The project identifies itself as a preview. Its current guardrail input model is text-focused, with other media described as future work. Execution failures do not trigger a tripwire by default, so teams must deliberately choose stricter error handling. Some checks may use third-party services, and paid OpenAI API calls can add cost.
How to use it well
Use it when a TypeScript application already calls OpenAI, Azure OpenAI, or the Agents SDK and needs policy checks around model traffic. Start with staged configuration, validate labeled datasets, inspect false positives and false negatives, then monitor tripwire and execution-failure results in production. It does not replace data-retention controls, legal compliance work, or broader application security.
Technical notes+
package.json publishes @openai/guardrails as compiled JavaScript plus declarations from dist, exposes the guardrails CLI, requires Node.js 18 or newer, and depends on openai, @openai/agents, and zod. src/client.ts subclasses OpenAI and AzureOpenAI, replacing chat and responses resources while retaining other client behavior. src/runtime.ts loads JSON configurations, validates check configuration with Zod, instantiates registry entries, and executes bundle checks through Promise.all; execution exceptions become failed results unless raiseGuardrailErrors is enabled. src/streaming.ts checks accumulated output every 100 text chunks by default and performs a final output check. src/cli.ts provides configuration validation, JSONL dataset validation, evaluation, and benchmarking. src/types.ts defines result and token-usage shapes, while src/registry.ts supplies the extensible in-memory specification registry.
Observed
- License
- MIT
- Primary language
- TypeScript, compiled to JavaScript with declaration files
- Package
- Public npm package named @openai/guardrails
- Interfaces
- Node.js library and guardrails command-line interface
- Platform support
- Node.js 18 or newer
- Client integrations
- OpenAI, Azure OpenAI, and OpenAI Agents SDK
- Evaluation data
- JSONL datasets with expected trigger labels
Read from README.md, Makefile, package.json, pyproject.toml, src/cli.ts, src/index.ts, src/spec.ts, src/types.ts, src/agents.ts, src/client.ts, src/runtime.ts, src/registry.ts, src/streaming.ts, src/exceptions.ts, src/base-client.ts.
What it can do
Validate LLM outputs against predefined schemas
LLM response and validation schema → Validated and structured data or validation errors
Filter harmful or inappropriate LLM content
LLM-generated text → Filtered safe content or content rejection notice
Enforce data type constraints on LLM responses
Raw LLM output and type specifications → Type-validated data or constraint violations
Retry LLM requests when validation fails
Failed validation results and retry configuration → New LLM request or final failure notification
Log safety violations and guardrail events
Guardrail execution data and logging configuration → Structured log entries and audit trails
Transform LLM outputs to required formats
Raw LLM response and target format specification → Formatted data matching required structure
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.