Agents and MCP

The zc-workstation MCP server, Claude Code plugin, skills and adapters for other agents.

.md

MCP server

zc-workstation mcp serves these tools over standard input and output:

status

Live state per target. Read-only.

list

Every declared item with its state, version, newer version and app, then what is installed but not declared. Takes state and kind filters. Read-only.

add

Declare a package or runtime in one target’s manifest, with its why and section. Edits the manifest only.

remove

Uninstall a package and drop it from the manifest. Destructive: only once the user okayed it.

capture

The manifest one target would declare to match this machine. Writes nothing. Read-only.

validate

Check the config offline. Read-only.

plan

Plan and save it. Returns a plan_id. Changes nothing.

apply

Apply a saved plan by plan_id. Refuses when live state moved.

doctor

Install health, security settings and AI agent configs. Read-only.

history

What was applied, newest first, and from where. Read-only.

drift

What the last daily drift check found. Read-only.

apply takes only the id of a plan the server saved, never a plan body. A plan that deletes is cancelled and the reply names the terminal command for the user to run.

Plans expire. A saved plan older than an hour is refused; the agent plans again and shows the user the new one.
Every call is logged. Each tool call, its arguments and whether it worked go to ~/.local/state/zc-workstation/mcp-audit.jsonl. An apply also lands in zc-workstation history as applied from an agent.
Tools say what they do. Each tool carries MCP annotations: read-only tools say so, add edits a file, remove is destructive, and apply changes the machine. A client can ask before the ones that write.
Resources. The server offers the config (zc-workstation://config), the last 50 applies (zc-workstation://history) and each target’s manifest (zc-workstation://manifest/<target>), so an agent reads them without a tool call.
Prompts. Each skill is also an MCP prompt, for clients without skills.

Claude Code plugin

The repository is a plugin: skills/, commands/, hooks/ and .mcp.json.

$ claude plugin marketplace add zoolcoder/zc-workstation
$ claude plugin install zc-workstation

setup installs the binary, writes a config and runs doctor.
plan-apply plans, explains, asks for a yes, and applies that plan by id.
add-software turns "install X" or "I need a tool for Y" into a manifest entry: it picks the package, writes it with add under a section with a why, then plans and applies after a yes. Nothing is installed by hand, so the next machine gets it too.
review reads doctor and list, then reports what is new in the tools you rely on, what is broken or drifted, and up to three ideas. It changes nothing until you pick.
Slash commands: /zc-workstation:plan, /zc-workstation:apply, /zc-workstation:doctor, /zc-workstation:add <tool or need> and /zc-workstation:review.
The PreToolUse hook asks before any apply with -yes or without a saved plan file.

What is new after an update

review reports features worth trying in the versions you moved to. It reads the release notes between the old and the new version, from GitHub releases when the homepage is on GitHub, and keeps only what touches each item’s why.

To know what an update moved, upgrade with zc-workstation upgrade. It saves the outdated list to ~/.local/state/zc-workstation/upgrades/, then runs apply -upgrade; review reads the newest file there.

$ zc-workstation upgrade             # everything outdated
$ zc-workstation upgrade -only brew:tmux

One setup for every agent

The agents provider writes the same MCP servers, rules and skills into Claude Code, Codex, Cursor and Gemini CLI, and pulls the Ollama models you name. A new machine gets them with the rest of apply.

Other agents

adapters/ holds the same rules for Codex, Cursor, Gemini CLI, Copilot and cloud agents.

Edit this page Last updated October 16, 2026