Know what your AI agents cost at API rates.
Per-seat billing gives you one number a month. contextbill reads the transcripts already on your disk and shows you the rest: which sessions ran expensive, which produced nothing, and how much loads before you type a word.
Every figure is API-equivalent — what that usage would cost metered at Anthropic's published API rates. On a subscription you paid a flat fee instead, so the dollars value the usage rather than restate your invoice.
Prefer the terminal? npx contextbill runs offline and uploads nothing.
How it works
Four steps, and your transcripts stay on your machine.
Point it at a folder
Usually ~/.claude/projects.
Parsing runs in your browser
Token counts, model ids and timestamps. Never message content.
Only totals are saved
Aggregates reach the server. Raw transcripts never leave the tab.
Compare over time
Reports are kept, so you can watch a number move after you change something.
The number nobody shows you
You pay for the same context on every turn.
Every session loads a system prompt, a tool catalog, connector definitions and instruction files before doing any work. That block is written to cache once, then re-read on every turn after it. A 500-turn session pays for it 500 times.
Privacy
What actually leaves your computer.
Aggregates only
Totals, per-model spend and session-level counts. No prompts, no completions, no file contents.
Paths are redacted
Folder names encode your OS username. contextbill strips that before anything is stored.
Or skip the account
The CLI does the same analysis with no network access at all. Same engine, same numbers.