egorfedorov CCO Tools — what your tools really cost
MCP and Agent token costs used to be constants picked once (mcp__* ≈ 200 in,
Agent ≈ 500). Real results vary by orders of magnitude — a "list all issues"
query and a one-row lookup are the same tool name and nowhere near the same
cost. On MCP-heavy sessions the budget meter was guessing at its biggest line
item.
CCO now measures every tool result and learns the average per tool name.
Run from the plugin root:
show (or no arguments)
node src/tool-costs.js
Prints the learned table: calls, average, worst case, and cumulative total per tool, sorted by total spend.
When reading it back to the user, the useful observations are:
- The top row by
totalis where their context budget actually goes. It is frequently an MCP server rather than file reads. - A row marked
·has fewer than 3 samples and is still using the built-in constant — its estimate isn't trustworthy yet. - A large gap between
avgandmaxmeans that tool is unpredictable; suggest narrowing its queries (filters, limits, pagination) rather than dropping it.
Pair with /cco-overhead mcp — that shows which servers are configured but
never called; this shows what the called ones actually cost.
reset
node src/tool-costs.js reset
Clears the learned table and returns every tool to its built-in constant. Useful after changing MCP servers or their configuration, when past measurements no longer describe the current setup. Confirm before running — the history isn't recoverable.