audit_usageLABSCO SUMMARY
The bill gets read for you, and the fix comes out as config.
The output that matters is the LiteLLM YAML: an audit that ends in a number is an argument, an audit that ends in a routing file is a change you can ship. The free tools cover the first pass - total, waste share, and what could be saved.
What it is
A cost-audit server for LLM spend. It ingests your invoice, classifies the spend, names the waste patterns it finds, and turns the fix into LiteLLM configuration. Three tools are free; one is on a paid tier.
What you get
- Total spend, waste share and the top three waste patterns from the invoice you supply
- Cheaper model alternatives for a model you name, with LiteLLM YAML for it
- A single saveable-per-month figure, with a confidence attached
- On the paid tier: a per-call breakdown, a thirty-day projection, ready-to-paste LiteLLM and Bifrost configuration, and prompt-rewrite recommendations
Requirements
Python and the package installed from the repository, plus your own usage or invoice data to feed it. It runs locally with no phone-home. get_pro_report is the paid tool; the other three work without paying.
Setup effort
One command — pip install git+https://github.com/miloantaeus/milo-cost-auditor-mcp.git
USE CASES
01
Finding out where the LLM bill actually goes
Feed it the invoice and get the spend classified, with the biggest waste patterns named rather than left to be inferred.
02
Getting a number to take to whoever asks
Turn the audit into one saveable-per-month figure with a confidence on it, next to the waste share it came from.
03
Moving a workload to a cheaper model
Ask what else could serve the model you are on, and see what the swap is worth before you make it.
04
Going from audit to a config you can ship
Take the routing suggestion as YAML, and on the paid tier the per-call breakdown that says which calls it changes.
4 tools
Listed in project docs. Your client may show a different list.
suggest_routingestimate_savingsget_pro_reportBEFORE YOU INSTALL
What to know first
Tools you get
4 tools named in the project's documentation.
Requires
mcp-cost-auditor on PATH
INSTALLATION
Connect your client
Server setup
TERMINAL
pip install git+https://github.com/miloantaeus/milo-cost-auditor-mcp.gitMCP Clients
Claude Code
CLAUDE CODE · JSON
{
"mcpServers": {
"milo-cost-auditor": {
"command": "mcp-cost-auditor",
"env": {
"MILO_COST_AUDITOR_PRO_KEY": ""
}
}
}
}Leave MILO_COST_AUDITOR_PRO_KEY empty to use the three free tools. It only unlocks the paid get_pro_report: paste the pro_key you receive after buying a plan (by email, or returned inline after a Lightning payment).
Claude Code — python -m
CLAUDE CODE — PYTHON -M · JSON
{
"mcpServers": {
"milo-cost-auditor": {
"command": "python",
"args": [
"-m",
"milo_cost_auditor"
]
}
}
}Other MCP Clients
Cursor, Continue and other MCP-aware tools: add a stdio server whose command is mcp-cost-auditor, in your client's own MCP config format. The README gives no per-client file for these.
