WSLでpyenv + Codex CLI + Serena MCPを使う：Codex CLIと組み合わせてAIコーディングを効率化する WSLでpyenv + Codex CLI + Serena MCPを使う：Codex CLIと組み合わせてAIコーディングを効率化する

When working with AI coding assistants on a local codebase, the simplest approach is to feed the entire contents of a file to the model. For small scripts this is fine, but as a project grows, the practice becomes inefficient and costly: the context window fills with irrelevant code, token usage rises, and the assistant struggles to locate the precise symbol it needs to change. A growing pattern combines pyenv, OpenAI's Codex CLI, and the Serena MCP server on Windows Subsystem for Linux (WSL) to give the AI a more structured, semantic view of a repository. The goal is to make AI-driven edits faster and more accurate without overwhelming the model.

WSL is the foundation here because it provides a real Linux environment on Windows, which suits command-line tooling and avoids many path and shell incompatibilities. Inside WSL, pyenv handles Python version management, letting you install and switch between interpreter versions per project. This matters because tools like Serena, which is written in Python, can have specific version requirements; isolating them with pyenv keeps a system Python untouched and reduces dependency conflicts. The typical flow is to install build prerequisites, clone pyenv, set a global or local Python version, then create a clean environment for the MCP server and any helpers.

Codex CLI is OpenAI's terminal-based coding agent. It runs in your working directory, reads and edits files, executes commands, and iterates on tasks from natural-language instructions. On its own it can read files directly, but it benefits from richer tooling exposed through the Model Context Protocol. MCP is an open standard that lets agents connect to external servers offering tools, resources, and prompts in a consistent way. Because Codex CLI supports MCP servers, you can extend it with capabilities it does not ship with, and Serena is one such server worth adding.

Serena is a coding toolkit that turns a repository into something the AI can navigate semantically rather than as plain text. It typically relies on Language Server Protocol backends to understand symbols, so the agent can find a function definition, list references, or edit a specific symbol without loading whole files. In practice this means the assistant retrieves only the relevant slices of code, which trims token consumption and tends to reduce mistakes on larger projects. Serena is usually launched through a runner such as uvx or a local install, then registered with Codex CLI as an MCP server so its tools appear during a session.

Wiring the pieces together generally follows a clear sequence. After preparing WSL and installing pyenv with a suitable Python version, you install Codex CLI and confirm it authenticates and runs. You then make Serena available, point it at your project, and add its server entry to the Codex CLI configuration. Once connected, the agent can call Serena's symbol search and editing tools instead of dumping files into context. The combination appears most useful in mid-to-large codebases where pinpointing the right location is otherwise tedious, and where careless full-file edits risk collateral changes.

It helps to understand the surrounding ecosystem. MCP has gained traction across multiple agents, including Claude's tooling, Cursor, and other editors, so Serena is not tied to one client. Alternatives and adjacent tools include Aider, Continue, and the various vendor CLIs, while serena-style semantic retrieval competes conceptually with embedding-based code search. Each approach trades off setup complexity against precision. Language-server-driven navigation like Serena's is likely to feel more deterministic than vector search, though it requires a working LSP for each language you target.

A few practical caveats are worth noting. Versions of Codex CLI, Serena, and the MCP spec evolve quickly, so configuration keys and command names may change; checking current documentation is advisable. Performance inside WSL is generally strong, but file-system boundaries between Windows and Linux can slow indexing if a project lives on the Windows side, so keeping the repository within the Linux file system usually helps. Finally, granting an agent the ability to edit and run commands carries obvious risk, and reviewing changes before committing remains prudent. Used carefully, this stack offers a reproducible way to give AI assistants the structure they need to work efficiently rather than blindly.