Give it a large, unfamiliar repository, and it will use search, grep, and shell commands to find exactly what needs to change. It doesn’t need a pre-built index; it uses the same tools you do.

It can use glob to understand the shape of your project, and grep to find the exact lines that matter. It reads only what it needs, keeping its context window focused and your token costs low.

Permissions categorise every action by risk. Reading and searching happens silently. More dangerous operations like editing files, running scripts, up to deploying and force pushing prompts you.

You can declaratively tune these rules in your config.toml or use the /permissions command to do it on the fly. Allow it to write freely in src/** but always prompt before touching terraform/. The defaults are safe, but you can tweak it however you want.

If a task requires a deep dive into a specific subsystem, mew can spin up a subagent to handle it. The subagent does the research or refactor and reports back, keeping the main thread focused and clean.

Context is precious. Instead of filling the main session with 50 file reads, a subagent does the dirty work, distills the findings into a short summary, and returns the result.

mew supports OpenAI-shape and Anthropic-shape API adapters out of the box. It comes with defaults for opencode-zen, opencode-go, z-ai, and deepseek, pulling model catalogs dynamically.

You can point it at any compatible endpoint, local or remote. Whether you want the intelligence of today's frontier models or the locality of LM Studio or Ollama, mew can adapt.

mew speaks standard agent protocols. It runs entirely in the terminal, or exposes a WebSocket daemon that browser and mobile frontends connect to. Your agent lives where you write code.

MCP (Model Context Protocol) servers are loaded automatically. If it speaks the protocol, mew can use it. Connect it to your database, your issue tracker, or your internal APIs without writing wrapper code.

Configuration lives in a simple config.tomlthat can either be edited with your favorite editor, or with the built-in TUI config tool. Credentials can be set via environment variables, the system keyring, or a credentials file. Everything is overridable.

Environment variables with the MEW_ prefix override config values. Drop a markdown file into .mew/skills/ to teach it a new trick, which instantly becomes a slash command.

Drop a markdown file in .mew/skills/ and it becomes a slash command. Share them with your team or keep them in your dotfiles. Skills are just prompt fragments and tool bundles.

For deeper integrations, mew exposes a JSON-RPC 2.0 plugin runtime over stdin/stdout with 18 hook points covering the full agent lifecycle. Plugins can register tools, slash commands, and hook into any stage of a turn.

Every session is saved locally as plain JSONL. You can read exactly what it did, what it thought, and what commands it ran. Nothing is locked away in a database — every session is a file you own.

You can tail -f a session if you want. It\u2019s just lines on disk. Subagent sessions reference their parent, so you can trace the full execution tree.

It’s not magic, it’s just a careful collaborator that runs on your machine, uses your tools, and respects your boundaries.