ccplug

Manage which Claude Code plugins and skills are enabled per project.

Claude Code installs plugins at user scope, so every project loads every plugin's skill descriptions at session start — a context/startup tax that adds up across a large workspace. ccplug keeps everything installed globally (auto-update on) and writes a per-project enabled subset into the settings cascade, so each project only pays for what it uses.

Built to be driven identically by a human or by Claude itself (ships a SKILL.md).

Install

cargo install ccplug

Or grab a prebuilt Linux binary (downloads the latest release tarball into /usr/local/bin, override with DEST=):

curl -fsSL https://raw.githubusercontent.com/sibincbaby/ccplug/main/install.sh | bash

Or from a checkout: cargo install --path .

The install.sh path also installs the agent-facing skill automatically. cargo install only installs the binary — run ccplug skill install once afterwards to drop the skill in (it's embedded in the binary; --force refreshes it).

As a Claude Code plugin

This repo is also a self-hosted plugin marketplace. The plugin ships the skill (not the binary), so install both:

cargo install ccplug                            # the CLI the skill drives
claude plugin marketplace add sibincbaby/ccplug
claude plugin install ccplug@ccplug             # the skill

Commands

ccplug list                  # every global plugin + its skills + enabled state + est cost
ccplug list --sort cost      # rank plugins by estimated always-on token cost (expensive first)
ccplug status                # what is EFFECTIVELY active in the cwd project after the cascade
ccplug enable  <target>...   # bulk enable
ccplug disable <target>...   # bulk disable
ccplug skill install         # write the bundled skill to ~/.claude/skills/ccplug (--force to refresh)

list shows a COST column — an estimate of each plugin's always-on token cost (its skill descriptions, which load every session) — with a footer totalling enabled vs all. status reports the project's enabled cost (enabledEst). Cost is a local chars/4 estimate; claude plugin details <name> gives exact numbers.

Deciding what to enable (for Claude, or you)

ccplug list --json returns each skill with its own description — the "what it does + when to use it" summary (the same text Claude uses to decide when to fire a skill). Decide by description, not by plugin name. A name like vercel says little; its skills' descriptions tell you whether this project will ever call them. Only open a skill's full SKILL.md body if a description is genuinely ambiguous — bodies are large and lazy-loaded, so reading them wholesale re-pays the very context tax ccplug removes.

Weigh fit × cost: a high-estTokens skill the project will never trigger is the first to disable; a cheap, occasionally-useful one can stay.

The right subset is dynamic — re-run list and adjust when the task moves into a new domain, a skill keeps not firing, or a newly-installed plugin appears. ccplug governs only the already-installed global set; discovering and installing a new plugin from a marketplace is a separate claude plugin install step, after which it shows up here like any other.

Targets

plugin            vercel                       whole plugin (bare name auto-resolves to name@marketplace)
plugin@market     vercel@claude-plugins-...    disambiguate a name shared across marketplaces
plugin:skill      vercel:nextjs                a single skill   (see "v1 limits")
plugin:*          vercel:*                     all skills of a plugin (v1: whole plugin)

Flags (all commands)

Flag	Meaning
--json	machine-readable output
--scope project|local|user	which settings file to read/write (default project)
--from FILE / --stdin	read targets as a JSON array of strings
--dry-run	print the change, write nothing

Scope → file: project = .claude/settings.json, local = .claude/settings.local.json, user = ~/.claude/settings.json. Nearer scope wins per key.

Examples

ccplug list --json
ccplug disable firebase pyright-lsp                 # trim a frontend project
ccplug enable --stdin <<< '["vercel","ponytail"]'   # agent / bulk path
ccplug disable vercel --dry-run                      # preview (warns it also drops the MCP server)
ccplug status --json

Restart the Claude Code session for changes to take effect.

v1 limits

Claude Code's native on/off switch is per-plugin (enabledPlugins). There is no native per-project toggle for an individual plugin-owned skill — skillOverrides only governs loose/project skills, not plugin skills.

So in v1:

• A plugin:skill target where the skill is plugin-owned returns {"ok": false, "reason": "unsupported-v1"} and does not fail the rest of the batch. Disable the whole plugin instead, or wait for v2.
• plugin:* toggles the whole plugin (with a warning).
• Disabling a plugin also removes any MCP server / LSP / agents it ships in that project; ccplug warns when so.

Per-skill extraction of plugin-owned skills (--extract + a sync step), an interactive pick, and project-type presets are planned for v2 — see docs/spec.md §9.

Safety

• Only enabledPlugins / skillOverrides keys are modified; every other settings key (and key order) is preserved — the file is never rewritten wholesale.
• The target file is backed up to <file>.bak before the first write.
• --dry-run shows the resulting file without writing.

License

MIT