Home›Practice
Practice
Practice
Hands-on guides and first-hand notes on agent frameworks, skills and AI tools.
Total 217
Lessons from Early MCP Adopters: Enterprise Case StudiesEnterprises use MCP to put several models behind one interface, add access control and audit logs, and connect internal data. This article groups early-adopter cases by pattern and notes which results are vendor-reported.MCP Protocol Features: Progress, Cancellation, and LoggingMCP servers can report progress, accept cancellation, expose parameterized resource URIs, and return consistent errors. Each feature is shown with Python code, and the 2026-07-28 changes are flagged.MCP Roots: Telling Servers Which Files Are RelevantMCP roots let a client tell a server which file locations are relevant. They are hints, not access control, and they are deprecated in MCP 2026-07-28. New projects should use tool parameters, resource URIs, or server configuration.MCP for real-time streaming: Kafka and Pulsar transportsMCP messages can travel over Kafka or Pulsar through custom transports. This piece walks through both skeletons, their blocking calls and version assumptions, and what to add before production.Multi-modal MCP: images, audio and generated images in toolsMCP tools can analyze images, transcribe audio and return generated images. This piece covers the parts of a multi-modal build, how the three samples are written, and which parts are still placeholders.Scaling MCP servers: horizontal, vertical and distributedScaling an MCP server combines horizontal instances, vertical tuning, resource optimization and distributed nodes. This piece explains each, shows how the Redis registration and heartbeat work, and flags fixes the samples need.Routing requests in MCP: content, load and tool routingMCP routing sends each request to a suitable model, server node or tool endpoint. It covers content-based routing, load balancing and tool routing, and flags the sample code's gaps before you copy it.MCP transports: writing a custom transport on Azure messagingMCP defines two standard transports. This piece explains when a custom transport on Azure Event Grid or Event Hubs makes sense, what the stateless 2026-07-28 rules require, and what it costs in interoperability.Practical MCP: building servers with the official SDKsSurvey the official MCP SDKs, then deploy a remote MCP server to Azure with azd and test it with MCP Inspector, with API Management in front for security.MCP Apps: letting tools return interactive interfacesLearn MCP Apps, where a tool returns an interactive UI as well as data: how the server, UI resource and IFrame fit together, and a working FAQ example.Configuring MCP servers in Claude Desktop, VS Code and CursorConfigure MCP servers in Claude Desktop, VS Code and Cursor: find each client's config file, add a server entry, restart, confirm the connection and fix common failures.Writing an MCP server with the stdio transportWrite a local MCP server that talks over stdio: the client launches it as a subprocess, then test it with Inspector and connect it to Claude Desktop.Your first MCP server, from zero to a running buildBuild a first MCP server with a calculator tool and a greeting resource in TypeScript or Python, then test it with MCP Inspector.Using Codex on WindowsUse Codex on Windows: install the desktop app, choose PowerShell or WSL2, pick a Windows sandbox mode, and fix common setup problems.Advanced Codex configuration: profiles, models and providersAdvanced Codex configuration: switch between profiles, override settings per run, connect custom or local model providers, and tune approvals and sandbox.Writing better tasks for CodexHow to write Codex tasks: name the result, give a reproduction recipe and constraints, plan before editing, steer mid-run, and ask for verification.Codex hooks: running scripts and MCP tools in the agent loopCodex hooks run your scripts or MCP tools at fixed points in the agent loop. Learn where hooks live, how trust works, and how to block a tool call.Running Codex in CI with the Codex GitHub ActionRun Codex in GitHub Actions with openai/codex-action: store the API key, choose permissions, and build a workflow that reviews pull requests.Using Codex in VS Code, Cursor and Other IDEsUse Codex inside VS Code, Cursor, Windsurf, Xcode or JetBrains IDEs. Covers installing the extension, opening the chat, adding editor context, reviewing changes and handing bigger tasks to the cloud.Codex Non-Interactive Mode: Scripting with codex execRun Codex from scripts and CI with codex exec. Learn to pass prompts, pipe input and output, choose sandbox permissions, get JSON or schema-bound results, and keep API keys out of shared job environments.Connecting MCP Servers to CodexConnect an MCP server to Codex with one command or a config.toml entry. Covers STDIO and HTTP servers, OAuth login, tool approval, timeouts and the settings that control which tools Codex can call.Codex Subagents: Running Agents in ParallelCodex can start several subagents in parallel and merge their results. Learn when this saves context, how to write a delegation prompt, pick models and reasoning levels, and define custom agents.How to Write AGENTS.md for CodexCodex reads AGENTS.md before starting any task. This guide shows how to split rules between global and project files, override them in subdirectories, add code review rules and confirm which files loaded.Troubleshooting Claude Code configuration that does not take effectWhen a CLAUDE.md, setting, hook, MCP server or skill does not take effect, check what actually loaded, rule out overrides, test in safe mode, and read the common error messages.