Session Baseline Overhead Audit

SkillDocs & knowledge

Audit the fixed context overhead every session starts with, system prompt, MCP tools, agents, CLAUDE.md, memory, measured from real transcript usage

Use Session Baseline Overhead Audit in Claude, ChatGPT or Ahel Desktop

Free. Sign in, add Session Baseline Overhead Audit and connect your AI. About a minute.

Also: Claude Code · Cursor · Codex

Then ask your AI: use the Session Baseline Overhead Audit skill

Details

Instructions available. Your AI can read the instructions. Execution depends on the setup they require.

Add Ahel to your AI once: Claude, ChatGPT, Cursor, Claude Code or Codex. Then ask it to use this.

Session Baseline Overhead AuditStart free

What this skill tells your AI

The instructions your AI receives, as published by egorfedorov/claude-context-optimizer in skills/cco-overhead/SKILL.md and read by Ahel’s review.

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:

  1. 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.
  2. Cost per session — what that baseline costs to write into the prompt cache each session.
  3. Itemization — the locally measurable parts (project + global CLAUDE.md, memory index, agent definitions) and the unattributed remainder (system prompt, tool schemas, MCP servers).
  4. 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.

Signals

GitHub stars
114
Forks
9
Last commit
Sep 2026
Advanced
Item type
skill
Key
cco-overhead
Source
github.com/egorfedorov/claude-context-optimizer