JSON Toolbox Blog
Schema · Structured Output · Function Calling · MCP · local tooling in practice
Compaction shrinks history; it does not replace tools JSON Schema.
As of 2026-10-30: on 2026-10-28 the first-party Claude API raised Messages and Token Counting request limits from 32 MB to 128 MB. Bytes …
Read the article → AnalysisAs of 2026-10-26: Claude on Google Cloud (Anthropic managed) shipped 2026-10-22. JSON Schema stays in output_config.format. Swap Base URL…
Read the article → AnalysisAs of 2026-10-24: OpenAI marked the Decisions API GA on 2026-10-22. Evidence goes in user-message input_json; answers live in the answers…
Read the article → AnalysisThe contract sits in output_config.format.schema. Sending output_format or temperature still returns 400.
Read the article → AnalysisRead stop_reason first. After end_turn, take the text block. minimum is outside the subset and returns 400.
Read the article → Analysisarguments is the tool input. output_text is the final JSON. requires_action is not the answer yet.
Read the article → Analysiscreated lists phases. status_ended carries result. completed can still hide a failed thread.
Read the article → AnalysisBedrock has no Structured Outputs. strict needs additionalProperties: false. A reply can start with an empty thinking block.
Read the article → AnalysisEditing tools drops the whole cache. The new contract sits in tool_addition. The MCP URL stays on mcp_servers. Pin the list with mcp_tool_listing.
Read the article → AnalysisinputSchema filters the subscription. payloadSchema constrains data. Delivery is a signed webhook. HTTP 2xx means accepted.
Read the article → AnalysisThe gateway only transcodes OpenAPI 3.x. tools/list is unauthenticated by default. Miss body, and the backend never sees the fields.
Read the article → AnalysisThe tool loop got longer, and the final reply has a Schema too. Live API is not on this card. If ids do not match, the result never comes back.
Read the article → AnalysisThirty thousand agents delegate to one another. Official guidance: treat the other agent's output as a claim. If the handoff fails Schema, the next agent should not reach.
Read the article → AnalysisHome MCP exposes five tools. The host only handles OAuth. If run_home_actions args fail Schema, the light should not move.
Read the article → AnalysisZero-trust judges intent. Schema judges shape. A legal object can still refund the wrong SKU. A missing field should never reach the engine.
Read the article → Analysisantigravity-preview-09-2026 defaults to Gemini 3.8 Flash. File edits are line-range patches. Search is grep_search. If args fail Schema, do not reach.
Read the article → TutorialPull toolCall.functionCalls from a Live session, validate args, then build FunctionResponse with the same id. Missing fields or a string treated as an object: do not reach.
Read the article → Analysisgemini-3.8-live defaults to async function calling and does not support Structured Output. Audio enters the session; a toolCall object comes out. Do not execute until Schema says yes.
Read the article → CompareNot a fork in the road. The harness runs the loop. MCP is one connection in tools. Functions come back to you; MCP calls are made for you. Both share one Schema.
Read the article → AnalysisThe 2026-09-10 public beta folded the Codex harness into a hosted session. You no longer own the loop. session, required_actions, and tool_result still need a contract you can validate.
Read the article → SecurityThe attack surface has moved from prompts to the tool bus and the Skill directory. This article unpacks malicious MCP servers, fake Skill handbooks, and how JSON Schema writes permissions as a contract you can validate.
Read the article → CompareA Skill is a handbook. MCP is the tool bus. Function Calling / API is the direct contract. Split SKILL.md, MCP servers, and JSON Tool Schema.
Read the article → GuideClaude Code, Codex, and Cursor Read first, then patch with exact strings. Workflows, strict JSON vs JSONC, and merging arrays.
Read the article → GuideUploading a file is only the entry. Split ChatGPT sandbox analysis and Work’s @Data, and how JSON / CSV enter the automatic analysis loop.
Read the article → TrendsFrom i-have-adhd and diagram-design to agentskills.io: how SKILL.md loads on demand, and the boundary with Rules / MCP.
Read the article → AnalysisSurfaces differ; the contract is a structured object. JSON, JSON Schema, the agent loop — plus one sample all three can line up on.
Read the article → AnalysisSiri AI acts through an App Intents Schema. Foundation Models’ @Generable compiles to JSON Schema. The agent still hands over an object.
Read the article → AnalysisHow GPU-hours, data centers, power, tokens, and inference add up — and what 7B / 70B / 405B / flagship each cost.
Read the article → AnalysisStructured Output is not “please return JSON.” Constrained decoding, and why GPT, Gemini, and Claude pushed Schema into the API.
Read the article → AnalysisThree vendors wrap the API differently; the contract layer is JSON Schema. Structured Output, Function Calling, and MCP.
Read the article → ComparisonThere is no single strongest model. Split GPT-5.6 / GPT-6 Astra, Claude Fable 5.1, Gemini 3.8 Flash, Grok 4.6, and Llama 4 by task.
Read the article → AnalysisGoogle has not announced a date. Pre-training was confirmed 2026-07-21. Official wording, the 3.x cadence, and Schema you can lock in now.
Read the article → TrendsResponses API, Programmatic Tool Calling, MCP, and the Agents SDK are already shipping. Direct vs programmatic paths, and validatable JSON.
Read the article → AnalysisTwo APIs, WordInfo JSON, word-level timestamps, speaker diarization — and why not to mix native STT with Structured Output.
Read the article → GuideInfer a Schema from real JSON. Lock types, required, and additionalProperties, then compare OpenAI, Claude, and Gemini wrappers.
Read the article →Go back to the Workbench, paste JSON or load a sample, and generate a Schema, run a Diff, or write Function Calling.
Open the Workbench →