MCP
Everything relevant to MCP servers: installation, usage, custom SKILL files, optimization, development with MCP servers
Trace initialization, verify the HTTP version header, and stop cleanly when client and server share no supported MCP protocol version.
Implement bidirectional MCP cancellation, stop costly work, suppress cancelled responses, and handle completion races without protocol errors.
Negotiate support, reread resources after update notifications, and remove subscription state when interest or the MCP session ends.
How MCP servers advertise URI templates, attach template-level metadata, and turn user or model input into concrete resource reads.
How MCP clients declare relevant filesystem roots, report changes, and rely on operating-system controls to enforce access.
How MCP servers request model completions through clients while clients retain control of credentials, models, context, permissions, and approval.
Negotiate sampling tool support, choose a tool mode, execute calls on the server, and return matched results without moving control to the model.
Use server/discover, ttlMs, cacheScope, and change notifications to cache MCP 2026-07-28 results without leaking private data.
Launch a local MCP server as a subprocess, pass credentials through its environment, and keep stdout clean for JSON-RPC messages.
Open subscriptions/listen, filter for supported changes, recover from stream closure, and refresh affected caches in MCP revision 2026-07-28.
How MCP tasks wrap long-running requests in durable state, expose polling metadata, and return results later through `tasks/result`.
Configure OAuth resource indicators, audience validation, and separate upstream credentials so an MCP server rejects misbound tokens and avoids passthrough.
How MCP behavior hints describe tool risk, and why approval policy must rely on trusted server identity and configuration.
Add outputSchema and structuredContent to an MCP tool, validate both sides, and evolve the contract without silently breaking clients.
Troubleshoot MCP 2026-07-28 migrations across discovery, request metadata, MRTR, subscriptions, deprecations, and legacy fallback.
How hostile instructions enter through MCP results, and why tool restriction, scoped credentials, approval, and complete call logs contain them.
Build a query-led, validated metadata schema through Cloudinary’s Structured Metadata MCP server without overloading uploaders.
How to run automatic tagging over a media library through an MCP server and keep the resulting tags searchable, trusted and dated.
What Cloudinary's Analysis MCP server exposes, why it bills as add-ons, why detection returns confidence scores, and why its endpoint is on the SSE path.
Pick MCP servers from the operations a project performs, not from the list available. Most projects need one; delivery-only projects need none.
A vendor CLI against an MCP server for bulk asset operations: setup, determinism, per-item token cost, and the cases where each one is the right tool.
Every connected MCP server loads its tool schemas into the context window before the first message, and a larger tool surface degrades selection accuracy.
Header-authenticated MCP client configs store a plaintext API secret on disk. How that file gets committed or logged, how to check, and what to rotate.
When a marketplace plugin is the right way to install MCP servers and skills, when a hand-written config entry is, and why most teams end up with both.