See what your LLM calls cost — then route to the cheapest model that still fits.

What you get

Know the spend. Then spend less.

ModelSpend sits in front of agents and code that already speak OpenAI-compatible APIs. Built because private projects still need a bill with project names on it.

• Cost per project and agent Tag with project / agent / labels. Report tokens by type, cost, and overhead from retries and escalations.
• Cheapest suitable route Hard constraints first, then the lowest-cost model with enough suitability evidence — inside your USD or token budget. Hard budgets refuse before the call.
• Shadow evals on your traffic Opt in: sample real requests on cheaper models in the background, compare answers, and project savings. Never stores prompts or answers — verdicts and reason codes only.
• Drop-in local server OpenAI-compatible /v1/chat/completions on localhost (or a small shared server with a token). Point any agent at it with model="auto".

Try shadow

Two minutes to a savings sample

Install editable, turn on a capped shadow sample (default $5 / month ceiling), use your tools as usual, then ask what you would save. Shadow stays off until you enable it.

git clone https://github.com/rmencke2/modelspend.git
cd modelspend
python3 -m venv .venv && source .venv/bin/activate
python -m pip install --upgrade pip && pip install -e .
modelspend shadow set --rate 0.05 --budget 5
modelspend shadow on
# ... use your tools through ModelSpend ...
modelspend savings

No keys yet? modelspend preview "Classify this ticket" --catalog synthetic --assume-keys still works for a dry routing preview.