GitHub
Docs Track agent cost

Track Claude Code and Codex cost per session

Track Claude Code and Codex token usage and cost per session, repo and model, locally. Estimates from tokens × list price are marked ~.

coding.sh reads the token counts that Claude Code, Codex and other coding agents write to their own transcripts, and turns them into tokens and cost per session, repo, agent and model. It runs on your machine. Nothing is sent anywhere unless you add a sink.

Get totals

Install it, import the sessions from before you installed, and ask for totals:

shell 
curl -fsSL https://coding.sh/install | sh
coding.sh ingest --backfill --since 30d
coding.sh usage --since 30d                  # by agent (the default)
coding.sh usage --since 30d --by repo
coding.sh usage --since 30d --by agent-model
coding.sh usage --since 7d --by day

Each row shows sessions, turns, total and output tokens, tokens per turn and the cache-read share. A COST column appears when at least one session in the window has a cost. All options: coding.sh usage.

Find the sessions that cost the most

shell 
coding.sh sessions --where 'sort:-cost limit:5'
coding.sh show <session-id>

The show header has a Tokens line with input, output, cache reads and cost. coding.sh ui opens the same session in the local viewer.

Where the dollar figure comes from

Transcripts carry tokens, not dollars. Each model call's cost comes from the first of these that exists:

  1. the bill from an LLM gateway, after coding.sh gateway enrich;
  2. a cost the agent reports: Claude Code headless runs (total_cost_usd), OpenCode, Pi, or an OpenTelemetry span with a cost attribute;
  3. an estimate: each kind of token times the model's list price.

Estimates are always marked:

WhereMark
usage~ before the figure
a call in showest
a session in show(estimated)
--jsonestimatedCostUsd

The price table covers current Claude, OpenAI and Gemini models and common open models, with long-context rates and the batch, flex, priority and fast tiers when the agent records them. A model with no price keeps no cost, never $0. usage names it and the fix: a [prices."<model>"] table in ~/.coding.sh/config.toml.

Which agents report tokens

AgentTokens
Claude Code, Codex, OpenCode, Pi, Factory Droid, Amp, T3 CodeEvery session
Antigravity CLI, Goose, GitHub Copilot CLIFrom their history files, imported by coding.sh ingest --backfill; their hooks carry none
Cursor CLINone, in its hooks or its stream output, so no cost either
coding.sh run -- <cmd>None, except Aider's token and cost lines in eval runs

What it does not do

  • It does not read your invoice. Unless a gateway or the agent reports a cost, the figure is an estimate at list price.
  • It does not know your plan. On a flat subscription, the estimate is what the same tokens would cost at API list prices, not what you paid.
  • It cannot add tokens an agent never wrote down.

Questions

Why does usage show ≥ in front of a cost?

Some sessions in that row have tokens but no price, so the figure counts only the priced ones. Below the table, usage says how many sessions have no cost and names up to five of their models.

Can I see cost from inside Claude Code?

Yes. coding.sh init registers an MCP server whose usage tool returns the same totals, with estimates marked. It takes the same where filter as --where.

Why are cache reads left out of tokens per turn?

Every model call re-reads the whole context, so cache reads grow with the number of calls, not the work done. TOK/TURN counts input, output and cache-write tokens of top-level sessions per user turn, leaving subagents out. The CACHE READ column shows the cache-read share.

Prices changed. Do old sessions update?

coding.sh usage --reprice recomputes estimated costs from the current price table first.

Next