Measure how many tokens every session of this project pays BEFORE any work happens — and where to cut.
Run:
node ${CLAUDE_PLUGIN_ROOT}/src/overhead.js
The report shows:
- Baseline — exact context size at the first assistant response (from the session transcript's API usage counts), latest and averaged over recent sessions, as a % of the working budget.
- Cost per session — what that baseline costs to write into the prompt cache each session.
- Itemization — the locally measurable parts (project + global CLAUDE.md, memory index, agent definitions) and the unattributed remainder (system prompt, tool schemas, MCP servers).
- Recommendations — what to trim and how (e.g.
/cco-claudemd, disabling unused MCP servers, pruning agent descriptions).
Then run the MCP usage audit — it turns 30 days of tracked tool calls into per-server verdicts and the EXACT removal command for servers that were never called:
node ${CLAUDE_PLUGIN_ROOT}/src/overhead.js mcp --servers <name1>,<name2>
Pass the MCP server names from your own tool list (mcp__<server>__*, one entry
per server). The script never reads ~/.claude.json — it holds account data —
so it only knows the servers you pass plus the project's ./.mcp.json.
Only if the report actually prints claude mcp remove ... commands, OFFER to
run them for the user (each removal repays in every future session; claude mcp add restores any time). Only run them after the user agrees.
If a server is listed as ? not observed, the tracker has no MCP data yet —
that is not a verdict. Never suggest removing those servers, and never
construct a claude mcp remove command the report did not print.
Present the output to the user as-is (it is already formatted). If the report says no transcripts were found, explain that the audit needs at least one completed exchange in a session for this project.
Key framing for the user: baseline overhead is paid in EVERY session, so a one-time trim repays itself continuously — it is usually the highest-leverage optimization available.