MCP & Claude Skills

Your agent gets better with tools (MCP servers) and packaged expertise (Claude Skills). byre’s job is to make both declare-once: say it a single time, and every session of the box – every rebuild, every worktree – comes up with the tool already wired and attributed.

MCP servers

byre mcp add github -- github-mcp-server stdio
byre mcp add tracker https://mcp.example.com/ --bearer TRACKER_TOKEN

A server is a URL (remote) or a command (run in the box). --global declares it for every project; --bearer NAME wires token auth without ever writing the token down – the header names an env var, and the value resolves inside the box at launch.

byre mcp list shows the effective set: what you declared, what your enabled skills contributed, and where each came from. Removing is symmetric – byre mcp remove <name> understands the difference between deleting your own declaration and switching off one a skill ships. A remote server’s network reach is attributed in byre status.

Authenticating a server

Declarations carry env var names, never token values – the declaration bakes into the image, and secrets don’t belong there. The value arrives at launch, from the box’s environment:

byre mcp add tracker https://mcp.example.com/ --bearer TRACKER_TOKEN

sends Authorization: Bearer <whatever $TRACKER_TOKEN is in the box>. Getting the token into the box is a separate, visible grant – [env_from_host] passes it through from your host environment at launch (recipe) – and command servers work the same way: --env NAMES declares what the process reads. Interactive OAuth flows aren’t wired; static tokens are the supported story.

Claude Skills

byre claude-skill add ~/claude-skills/tdd-loop

A Claude Skill is a directory with a SKILL.md at its root. Declare it once and it loads in every session as /tdd-loop – no per-box setup, no copying into the project. --global shares it across projects; byre claude-skill list shows the effective set with the same attribution as MCP. Skills you build inside the box win over byre-delivered ones of the same name – your working copy shadows the declaration while you iterate on it.

Where declarations live

Both ride the config cascade, so the right scope is one decision: this project, your global defaults, or a shared layer. Skills you enable can ship their own declarations, and you can switch any single one off without disabling the skill. The precise wiring contract – field shapes, header templating, closure semantics – is in the packaging reference.