LLM Tool Call Debugger
Inspect function-call arguments, schemas and call/result linkage locally for OpenAI, Anthropic and Gemini.
Inspect custom function calls as inert data. No tool, command, URL or prompt is executed. Validation covers the fields and schema subset documented below; it is not a provider acceptance guarantee.
Ready. Paste payloads or load a synthetic sample.
Argument subset: object, array, string, number, integer, boolean, null/nullable union, required, enum, minimum/maximum and boolean additionalProperties. No references, patterns, formats or composition keywords. Unsupported schemas and declared __proto__ properties fail closed. Names outside the portable 1–64 ASCII identifier profile are flagged for provider review.
Responses API rule source · Local adapter 2026-10-16.1. Provider schemas can change; no documentation is fetched while validating.
Enable JavaScript to validate locally.
Inspect custom function lifecycles.
Choose OpenAI Responses, Anthropic Messages or Gemini Interactions and paste tool definitions, model calls and results. Calls are normalized locally before validating schemas and linkage. Legacy Chat Completions, generateContent, built-in/server tools and streaming deltas are outside this tool’s scope.
Keep provider boundaries explicit.
OpenAI arguments are JSON-encoded strings and results link through call_id. Anthropic inputs are objects and results link through tool_use_id. Complete Anthropic history additionally checks assistant/user roles, adjacent turns and result blocks before normal content. Gemini call id links to result call_id; an optional result name is compared when present.
Understand validation limits.
The supported JSON Schema subset reuses the existing schema analyzer and Zod instance validators. References, regex patterns, formats and composition are rejected rather than ignored. Declared __proto__ properties are unsupported. Schemas allow eight levels and 100 nodes across all definitions. Extra properties are rejected only when additionalProperties is false. OpenAI strict declarations are checked explicitly. Nested media results receive a scope notice; semantic correctness, permissions, retry behavior and actual execution are never verified.
Keep payloads local.
No function, URL, shell command or tool is executed. Diagnostics show indexed paths and fixed explanations instead of argument contents. Reset and leaving the page clear content. Only synthetic samples are supplied.