{ "issuer": "pinned" }
How an MCP client pins its OAuth issuer
Python SDK 1.30 or 2.2 is not the whole fix. Machine-to-machine still needs issuer, and old client_info has to go.
Read moreJSONVue · Developer notes
Blog
JSON tips, step-by-step tutorials, comparisons, and best practices to help you understand and work with JSON in more depth.
- Articles
- 33
- Topics
- Tutorial
Earlier posts
Previous tutorials
-
Hand an existing REST API to an agent: declare MCP tools in OpenAPI
The gateway turns tools/call back into the original HTTP request. The body sits under body, and an API key cannot lock tool discovery.
-
How to validate an A2A 1.0 Card: green Ajv, then the spec table
Official a2a.json is non-normative and often skips required. Point at AgentCard, assert §4.4 non-empty, then switch the pointer for Task.
-
How one agent finds another: well-known, catalog, then the Card
A2A does not specify a catalog API. Three strategies land on the same Card. Only then do you select, pick an interface, and send a Task.
-
A2A Agent Card 1.0: name, interfaces, and skills.tags
Fill §4.4 field by field. Top-level url is 0.3. No tags, no search.
-
Google Agent Registry 2026: the Card is source, the catalog is search
Not a third protocol. The Registry indexes A2A Cards and MCP toolspecs. Searchable first, then delegate.
-
From Tool Calling to Skills + Plugins: why the 2026 architecture changed
The call hop did not retire. The full menu at startup did. Discovery thins the window; packaging stops the fork.
-
Plugin Manifest in practice: how an agent grows a capability after install
Not a fine-tune. Five hops later, memory holds identity, skill metadata, and tools/list. Check it with a snapshot JSON.
-
Skills vs MCP vs Plugins: which layer in 2026
Not a three-way vote. Brief, hand, and box each own a discovery hop. Walk four questions before you pack.
-
Google Agent Plugins 2026: Skills, MCP, and JSON Schema
A plugin is a box, not another tool protocol. Read the closed schemas for plugin.json and mcp.json, and why the Cloud Developer Plugin ships that way.
-
Generate JSON API docs, TypeScript types, and a client from OpenAPI
One OpenAPI 3.1 schema grows docs, types, and a typed client. Check the contract first, then run codegen—do not hand-write three copies.
-
How to deploy a remote MCP server to production: stateless MCP + HTTP load balancer
After 2026-07-28, stop sticking sessions. JSON-RPC crosses a normal LB, SSE buffering stays off, rollouts drain long streams first.
-
Add OAuth 2.1 to an MCP server: Authorization Server to Access Token
Remote MCP 401s, RFC 9728 discovery, PKCE for a token, Bearer on tools/call—auth lives on the HTTP header, not in the JSON-RPC envelope.
-
OpenAI DevDay 2026 developer guide: four API tracks and a JSON Schema list
What may change in Responses, Structured Outputs, Tool Calling, and MCP—and six schemas to freeze now. Predictions are not the agenda; contracts can be ready anyway.
-
What is AI Agent State? JSON state, memory, and task execution
An engineering definition of 2026 agent state: the checkpoint is the position, memory is only recall, who writes the JSON in a workflow vs a loop, and how run/step become recoverable.
-
OmniRoute: Connect 352 AI Providers Through One API
Coding agents enter the multi-model era: how a local /v1 gateway unifies Claude Code, Cursor, and Codex, then reroutes requests when quotas run out or providers go down.
-
Will Siri AI become an AI agent? App Actions and tool calling
Siri in 2026 can pick an App, fill arguments, and read the screen. Map that to an agent definition, then split Apple Intelligence, App Intents / App Actions, and the two parallel tool contracts.
-
Apple Intelligence vs ChatGPT vs Gemini: 2026 comparison
Keep product, model, and API on separate layers: what Siri AI and PCC, the ChatGPT extension, and the Gemini foundation each own — and why the three JSON contracts must not be mixed.
-
What is MCP? JSON-RPC, AI agents, and tool calling
An engineering definition of MCP: which layer the protocol, the JSON-RPC envelope, the agent loop, and model tool calling each own — and why arguments still need a local check.
-
What is a 1M token window? GPT vs Claude vs Gemini
Turn 1M tokens from a marketing number into a window budget: what fits, how output and pricing fork, and why long JSON should be validated locally first.
-
What is an AI agent? Tool calling, function calling, and JSON
An engineering definition of the 2026 agent: the observe–decide–call loop, how two product names map, and how three JSON payloads share one Schema.
-
Why ChatGPT, Claude, and Grok went down together
On 3 September 2026 three front-door models overlapped. Separate the official timeline, Azure ingress clues, and retry cascades — and how an Agent JSON pipeline should fail over.
-
Why Apple is leaning harder on Gemini in 2026
January’s joint statement put the next AFM on Gemini. June expanded PCC onto Google Cloud. Separate official facts from reported numbers—and why your JSON contract did not change owners.
-
OpenAI DevDay 2026: 10 AI API features to watch
Before 29 September, track Responses, MCP, Structured Output, and enterprise identity from the changelog—predictions are not the agenda; JSON contracts can be ready anyway.
-
A2A vs MCP 2026: JSON, tools, and agents
MCP connects tools; A2A connects agents—complementary, not rivals. Where JSON fits in Agent Cards, tools/call, and task lifecycles.
-
MCP and JSON Schema: 2026 agent tool calling data structures
Align inputSchema, function.parameters, and Structured Output; full flow from model arguments to tools/call plus validation.
-
Why AI JSON still fails: 2026 Structured Output guide
Structured Output locks shape, not semantics. Four error layers, 2026 API map, Schema rules, parse→Schema→business pipeline.
-
OpenAI Assistants API sunset August 26, 2026: migrate to Responses API
Assistants, Threads, Runs go away; map to Responses + Conversations, backfill Threads manually, reuse vector stores.
-
DeepSeek V4-Flash: API, agents, tool calling, and JSON output changes
V4-Flash public beta: model migration, Thinking replay of reasoning_content, Tool Calling strict beta, json_object output, and empty-response handling.
-
What is Stateless MCP? 2026 stateless architecture, JSON-RPC, and Remote Server explained
The 2026-07-28 MCP spec makes the protocol layer stateless: every JSON-RPC request carries its own version and capabilities, and Remote Servers can run behind ordinary HTTP load balancers.
-
How does an Apple AI agent connect apps, APIs, and tools? What role does JSON play?
The system agent uses App Intents; the in-app model uses Foundation Models tools. Split the three call chains, and see what JSON actually does in arguments, schemas, and HTTP bodies.
-
How does the Gemini API output JSON? A complete Structured Output and JSON Schema tutorial
Constrain Gemini’s response shape with Structured Output and JSON Schema instead of begging the prompt for “JSON only”. Covers JSON mode vs Schema mode, Python/JS samples, and production validation.