Which APIs are supported?

The OpenAI Chat Completions format (model + messages array with system/user/assistant roles) and the Anthropic Messages format (model + max_tokens + messages with user/assistant roles and an optional top-level system field).

Does it send my payload to the provider?

No. The inspector only parses and validates the JSON locally in your browser. It never makes a network request and never transmits or stores your payload.

How accurate is the token estimate?

It is a heuristic — roughly one token per four characters of message content — so treat it as an order-of-magnitude guide, not an exact count. Use a dedicated tokenizer for billing-grade numbers.

What does a warning versus an error mean?

An error means the request will almost certainly be rejected (missing required field, invalid role, out-of-range value). A warning flags something unusual but legal — for example a very high temperature or an empty message — that you may have set by mistake.

LLM API Payload Inspector

Catch a bad LLM request before the API does

A single malformed field — a missing max_tokens on Anthropic, a stray system role in the OpenAI messages array, a temperature of 5 — turns into a 400 and a wasted round-trip. This inspector parses your full request body and validates it against the provider’s rules locally, so you fix the payload before you ever hit the network.

How it works

You paste the JSON body and pick the provider. The inspector first parses the JSON and surfaces any syntax error precisely. Then it runs format-specific checks: that model is present and a string; that messages is a non-empty array with valid roles in a sensible order; that Anthropic requests include a positive integer max_tokens and put the system prompt in the top-level system field rather than the messages array; and that sampling parameters (temperature, top_p, top_k, presence_penalty, frequency_penalty) sit inside their legal ranges. Finally it sums the character length of all message content for a rough token estimate.

Tips and notes

Each finding is tagged pass, warn or fail. Fails are blocking — the API will reject the request — so clear those first. Warnings are legal but suspicious (an empty message, an unusually high temperature, both temperature and top_p set together) and worth a second look. The token estimate is deliberately conservative and approximate; if you are close to a context limit, verify with a real tokenizer. Because everything runs in the browser, you can safely paste payloads that contain prompt text without it leaving your machine.