Home›Practice
Practice
Practice
Hands-on guides and first-hand notes on agent frameworks, skills and AI tools.
Total 217
Deploying MCP serversAn MCP server can run on your own machine or be deployed as a container app in the cloud. This guide covers the local steps, the Azure Container Apps sample, and why SSE endpoints are now legacy.Testing MCP serversThere are three ways to test an MCP server: MCP Inspector for visual and CLI checks, curl for raw HTTP calls, and unit tests for each tool. This guide shows how to run each one.Debugging MCP servers with MCP InspectorUse the MCP Inspector web interface to test an MCP server's tools, resources and prompts, read its protocol message log, and troubleshoot connection and tool problems.MCP over Streamable HTTPRun an MCP server over Streamable HTTP: check the spec version first, then learn notifications and progress updates, security basics and how to migrate from SSE.Using MCP servers in VS CodeConnect an MCP server to VS Code and GitHub Copilot Chat: configure .vscode/mcp.json, start the server, tick the tools you want, and call them with a plain-language prompt.Connecting an LLM to MCP toolsAdd an LLM to an MCP client: convert the server's tools into the format the model expects, then let users ask in plain language and run the tools the model picks.Building an MCP clientWrite a small MCP client that starts an MCP server over stdio, lists its tools, resources and prompts, and calls them from code in TypeScript or Python.Deploying an MCP server over HTTP with FastMCPRun a FastMCP server over HTTP: direct and ASGI deployment, custom paths, authentication, multiple workers with stateless mode, and checks before going live.Resources and prompts in FastMCPHow to expose data with @mcp.resource, generate resources from URI templates, mask error details, and define reusable prompts with @mcp.prompt.Defining tools in FastMCPHow to define tools in FastMCP: parameter types and defaults, validation modes, docstring descriptions, return values with structured output, and ToolError handling.FastMCP quickstartInstall FastMCP, write a server with one tool, run it over stdio or HTTP, call it from a client, and learn the basics of deploying it.Writing an MCP server with the official Python SDKInstall the official MCP Python SDK, write a server with one tool and one resource, test it in the MCP Inspector, run it over HTTP and call it from a client.Running Codex with OllamaRun Codex CLI on Ollama models: start it with one ollama launch command, switch models inside the session, or keep the setup in a profile file. Covers install, the 64k context note and web search.Writing and calling Codex skillsA Codex skill is a folder built around a SKILL.md file that packages a reusable workflow. This guide covers how to write one, where Codex looks for it, how to invoke or disable it, and how skills differ from plugins.Codex sandbox and approval modesCodex sandbox limits what commands can touch, while approval policy sets when the agent must stop and ask. This guide covers the three sandbox modes, two approval policies, common combinations and Linux setup.Configuring Codex with config.tomlCodex reads its settings from config.toml at user, project and system level. This guide explains how the layers combine, and how to set the model, approval policy, sandbox, web search and feature flags.Signing in to Codex CLI with a ChatGPT account or an API keyCodex CLI offers two sign-in methods: a ChatGPT account for subscription access, or an API key for usage-based billing. This guide covers the login, status and logout commands, credential storage, and headless fallbacks.Running Claude Code against Ollama modelsConnect Claude Code to local or cloud models through Ollama's Anthropic-compatible API. Covers one-command launch, manual environment setup, context length, and tool-calling limits.Claude Code in large codebasesIn a large repository, nested CLAUDE.md files, deny rules for generated code, sparse worktrees, and code-intelligence plugins keep Claude focused on the code a task touches. Covers where to start, what to block, and how to grant access.Scheduled tasks and routines in Claude CodeCloud scheduled tasks run on Anthropic's infrastructure even when your laptop is closed. Session-scoped /loop tasks poll while your session stays open. This guide covers both and their limits.Managing Claude Code sessions and contextResume, name, and branch Claude Code sessions, then keep the context window under control with /clear, /compact, and the rules for what survives compaction.Parallel Claude Code sessions with git worktreesRun several Claude Code sessions in parallel, each in its own git worktree with its own files and branch. Covers startup, environment setup, cleanup, and where isolation stops.Claude Code plugins: install, build and publishWhat a Claude Code plugin contains, how to install one from a marketplace, build a first plugin with a skill, validate and load it, and share it or check its scope.Claude Code GitHub ActionsSet up the Claude Code GitHub Action so @claude can review pull requests and work on issues, choose interactive or automation mode, and control cost and access.