llmledger

llmledger-mcp

Ask Claude (or Cursor, ChatGPT desktop, any app that speaks MCP) which AI model to use and what it costs, in the middle of your work. The answers come from llmledger: LiteLLM's public price catalogue, Arena's ratings by task, and every price change since 2023 with how likely a price is to change next. Rebuilt every day.

Install

Needs uv. Nothing to host and no keys: the plug-in runs on your computer and reads the day's data from the llmledger page, cached for a day, with a bundled copy for offline use.

Claude Code:

claude mcp add llmledger -- uvx llmledger-mcp

Claude Desktop (Settings, Developer, Edit config) or Cursor (.cursor/mcp.json):

{
  "mcpServers": {
    "llmledger": { "command": "uvx", "args": ["llmledger-mcp"] }
  }
}

A LiteLLM proxy can offer it to a whole team through its MCP gateway, as a stdio server with the same command.

The same answers also come as an Agent Skill, which takes far less of an agent's context until it is used and needs only Python: the folder skills/llmledger in the repository, https://github.com/pouriaetab/llmledger/tree/main/skills/llmledger In Claude Code: /plugin marketplace add pouriaetab/llmledger, then /plugin install llmledger@llmledger.

What you can ask

Tool	For example
suggest_for_project	"Which model and effort for this project?" with the starting prompt; pay can be claude or chatgpt to compare within a plan
compare_models	"Opus 5.5 or Fable 5.1 for coding, and when is Fable cheaper?"
estimate_cost	"What do 2,000 calls a day of 10K tokens in and 1K out cost on these three?", and with cached_tokens and cache_write_tokens, how much the cache saves
find_models	"The best rated coding models under $10 per 1M answer tokens that read images"; or name to look one model up with every host that lists it
price_history	"Has Claude Opus 5.5's price changed, and how likely is it to change soon?"; with date, "What did it cost on 1 March?" Each change links the LiteLLM change behind it
not_yet_priced	"Which new models on Arena have no price yet, and how long does a price usually take?"
estimate_time	"How long will 5,000 calls of 2K tokens in and 600 out take on Opus 5.5 through the API, 4 at a time?" Learned from your own runs
log_run	"I timed this call: 2,100 in, 640 out, 9.2 seconds", or a promptfoo results file; kept on your computer only
forecast_job	"Here is a promptfoo run of 20 of my 5,000 tests: what will the whole job cost and take on each model, and which should I pick?" Reads the file only
local_or_cloud	"Should I buy a $2,000 computer to run llama3.1:8b at home, or keep paying for GPT-6.1 at 30,000 calls a month?" The local model's speed from your own runs, electricity from EIA or your price
price_my_usage	"Here is my ccusage session --json (or codeburn export -f json) file: what would last month's work have cost on every other model, which ones keep the quality for less, and is the gap real?" Each session at the price of its day, Arena ratings for that kind of work, a 95% range per session; token counts only

Plus a resource, llmledger://snapshot (data dates and sources), and a prompt, pick_a_model.

Good to know

• Costs and ratings are estimates. Arena ratings say which answers people preferred, not what will work best for you. Plan use (Claude, ChatGPT) assumes a plan runs down in step with API prices, because plans do not publish per-model limits. Not financial or procurement advice.
• Effort levels come from LiteLLM; the suggested level is a rule of thumb. No open source measures tokens per task at each effort level yet.
• Many price changes in the catalogue fix a wrong entry rather than change a price: in a labelled sample, about 6 in 10 first price changes. price_history says what each flag on a change means (a likely correction, a unit fix, a change undone) and how often such changes were corrections.
• Every answer carries the date of its data. When the data is 3 or more days old, the answer says so.
• price_my_usage prices other models at your own token counts; another model tokenises the same work a little differently (usually within 10 to 20%) and may answer at another length. Models you used are compared on the sessions each one got, which were not the same work, and the answer says so. A model with no cache price has cache tokens counted at its input price, and the row says so.
• Times come only from your own runs (and the sample's own timed calls for forecast_job), always as a usual time and a slow day. How often a model was right is the share of a sample's checks it passed, with a 95% range; with 20 tests the range is wide, and the answer says when two models cannot be told apart.
• What the plug-in reads: a prompt given to suggest_for_project is only counted and matched against word lists; a promptfoo results file (log_run, forecast_job) is read for numbers and names only (provider id, latency, token counts, pass or fail), and the prompts, answers and grader text inside it are never passed on or treated as instructions; a ccusage or codeburn export (price_my_usage) is read for model names, dates, session ids and token counts only, never project paths or file names. Timed runs stay on your computer as numbers. The only network call is the daily download of data.json.
• Set LLMLEDGER_DATA to a local data.json or another address to use your own copy.

Data and licences

Prices: LiteLLM's model_prices_and_context_window.json (MIT). Ratings: the Arena leaderboard (CC BY 4.0). Electricity: the U.S. Energy Information Administration's Electric Power Monthly (public domain). Price history and outlook: llmledger (MIT). Built with FastMCP (Apache-2.0). Independent; not affiliated with BerriAI, LiteLLM or Arena.