AI & LLM
Tools for building and debugging LLM applications.
- JSON Schema BuilderBuild a schema visually and export it as a tool definition.
- Structured Output ValidatorCheck model output against a schema, with the path that broke.
- Prompt DiffSee which instructions changed, and what the change costs.
- MCP Config ValidatorCatch the mistakes that break an MCP config in practice.
- RAG Chunking PlaygroundSee the chunks your settings actually produce, before you index.
When you would reach for these
Almost every integration between an application and a language model fails in the same place: the boundary where free-form generated text has to become a value your code can rely on. The model returns something that looks right, your parser disagrees, and the failure surfaces three layers away from where it started. The tools here exist for that boundary.
Start with the Structured Output Validator when a response will not parse or will not validate. It separates the two cases, which matter differently — a JSON syntax error usually means truncation or a stray fence, while a schema violation usually means the model answered a question your schema did not quite ask. The JSON Schema Builder is the other half of that loop: a schema with vague field names and everything optional gives a model very little to aim at, and tightening it is often a faster fix than rewriting the prompt.
Prompt Diff is for the case where a prompt stopped behaving and nobody can see why. Prompts accumulate trailing whitespace, smart quotes pasted from a document, and invisible characters that survive a copy between editors. None of that is visible by eye, and all of it changes what the model receives.
The MCP Config Validator and the RAG Chunking Playground cover the two configuration-shaped problems: a server that will not connect, which is usually one missing or misspelled field, and a retrieval step returning too much or too little, which is usually chunk size and overlap rather than the model.
Why these run in your browser
Prompts and model responses are among the most sensitive text a developer handles. A prompt often carries a system instruction you would not publish, and a response often carries whatever customer record was passed in as context. Pasting that into a remote service to get a validation error is a poor exchange.
Every tool in this category processes entirely on your own device. Your tool input is processed locally in your browser and is not intentionally uploaded to our servers. Advertising and analytics providers may still process normal page, device, cookie and network information.