Unclog — Security skill for Claude Code

unclog

[](https://pypi.org/project/unclog/) [](https://github.com/thomaschill/unclog)

**Every agent, skill, slash command, and MCP server you've installed occupies your Claude Code context window on every turn.** `unclog` scans your install, shows you what's costing tokens, flags MCP servers you haven't called in 30 days, and lets you handpick what to delete — locally, with no telemetry, no accounts, and no network calls.

# install
uv tool install unclog

# run
unclog


What it shows

• Every agent in ~/.claude/agents/, with its token cost.
• Every skill in ~/.claude/skills/, with its token cost.
• Every slash command in ~/.claude/commands/, with its token cost.
• Every MCP server declared in ~/.claude.json (global and per-project), with its invocation count over the last 30 days so you can see which ones are actually pulling their weight. Servers with zero invocations get an [unused] flag.
• A baseline — the total tokens loaded before you type a single message.

Why no per-MCP token count?

Modern Claude Code session logs no longer record the tools schema, so unclog cannot recover an exact token cost for each MCP server without spinning up the server and querying it. We've chosen to stay lightweight (no network, no spawning user-configured commands) and show `— tok` honestly rather than guess. **Invocation count is the next-best signal**: a server you haven't called in 30 days is paying schema rent for nothing.

Invocation counts come from walking every session JSONL under `~/.claude/projects/` (parent and subagent files) and tallying `tool_use` blocks per server. Counts are per server *name* — if the same MCP is declared at both global and project scope, both rows show t