Prompt Inspector

Detection runs as a five-layer funnel, and each layer filters a little finer than the last. A global hash cache returns a prior result for a repeated input with near-zero latency. A sensitive word layer matches keywords and regex patterns using an Aho-Corasick automaton and short-circuits on a hit. A semantic layer uses vector embeddings to catch paraphrased attacks and obfuscated injections that keyword rules miss. An optional AI review handles gray-zone inputs, and an arbitration step aggregates the layers into a final verdict with a confidence score.

Most requests resolve in the first two layers in under 3ms, according to the project. Full semantic analysis completes in about 30ms. The funnel design is what keeps those numbers down: an obvious threat is stopped at the keyword layer, and only ambiguous inputs travel further down the pipeline. That split lets the engine stay fast on the common case while still spending real compute on the hard ones.

Each tenant defines its own word lists, and the engine supports two matching strategies. Literal matching runs on an Aho-Corasick automaton, which scans the whole input in a single pass with linear time no matter how many keywords you configure. Regex patterns cover structured data such as phone numbers and card formats. When a word matches, the pipeline short-circuits, so the layer works as both a safety net and a speed shortcut.

Inputs that land in an ambiguous middle ground are escalated to a language model for a second opinion. Supported reviewers include DeepSeek, Qwen, and Gemini. This layer exists to cut false positives, and it stays optional so a deployment can keep the pipeline fully deterministic when it needs to. A clear score from the semantic layer skips the review entirely.

When the AI reviewer confirms a new attack, the engine generates variant payloads and adds them to the vector database on its own. Over time the payload library grows from the threats the system actually sees, not from a fixed list someone maintains by hand. The enterprise edition adds a maintained payload library at million scale with real-time updates drawn from production threats.

Prompt Inspector ships a REST API, an official Python SDK, an official Node.js SDK, and an MCP server. The MCP server works with Cursor, VS Code Copilot, Claude Desktop, and Dify, and a Skill integration covers OpenClaw and Claude Code. Every client authenticates with the same App API Key, sent through an X-App-Key header or a bearer token, so one key covers the whole stack.

The core is open source under the AGPL-3.0 license and can be cloned and run on your own infrastructure. The enterprise edition layers on payload library maintenance, customization, private deployment, and dedicated support. That split gives a solo developer a working guardrail for free and gives a company a supported path when it needs one.