promptfoo
Open-source LLM eval and red teaming: test prompts, compare models, catch jailbreaks and data leaks locally and in CI.
🔗 Visit promptfoo
Description
When companies build apps on top of AI models — chatbots, assistants, support tools — they need to check that the AI answers correctly and cannot be tricked into revealing or saying things it shouldn't. promptfoo is a free tool that runs those checks automatically: part spell-checker, part security guard for AI applications.
promptfoo is an open-source platform for testing and securing LLM applications. Its evaluation side lets developers define test cases for prompts, agents and RAG pipelines, then compare outputs side by side across providers — OpenAI, Anthropic, Azure, Bedrock, Ollama and others — from a CLI or a local web UI with live reload. Its security side automates red teaming: generated attacks probe for prompt injections, jailbreaks, PII leaks and business-rule violations, and real-time guardrails plus an MCP proxy defend agent traffic in production. Everything runs locally by design — prompts never leave the machine during evaluation — and the same checks wire into CI/CD so regressions and vulnerabilities block a pull request instead of reaching users. The project is MIT-licensed TypeScript with over 23,000 GitHub stars and a 300k+ user community feeding its threat intelligence. promptfoo was acquired by OpenAI in March 2026 but remains open source and independently maintained, with an enterprise tier for larger teams.
💬 Our review
The short version: if you build anything with AI and want to test it seriously without paying for a platform, this is the free tool most people start with — and rightly so.
promptfoo earned its position as the default open-source LLM eval tool honestly: config-as-code test cases, every major provider supported, and a local-first design that security teams can actually approve. The red-teaming module is the standout — it turns adversarial testing from a consulting engagement into a CI job. The March 2026 OpenAI acquisition cuts both ways: more resources, but keep an eye on neutrality when the tool scores OpenAI models against competitors — the MIT license and open codebase are the real safeguard, and forks are viable if governance drifts. Compared to alternatives: DeepEval offers a richer metrics library but less tooling around it; Braintrust adds a polished paid platform (UI, datasets, gates) that promptfoo doesn't try to match; Langfuse covers the observability side promptfoo ignores. For CI-based evals and red teaming the free tier is a no-brainer; you only pay if you need the enterprise layer.
💰 Pricing
📊 Global score
🤖 AI-enriched data
MIT open source, free for local and CI use; enterprise tier on request
Pros
Free red teaming + evals in CI
Local-first: prompts never leave your machine
Works with every major model provider
Cons
No team UI without the enterprise tier
Config-driven workflow has a learning curve