https://github.com/mbailey/ccexplorer
Cost Explorer for Claude Code
https://github.com/mbailey/ccexplorer
analysis claude-code cost
Last synced: 2 days ago
JSON representation
Cost Explorer for Claude Code
- Host: GitHub
- URL: https://github.com/mbailey/ccexplorer
- Owner: mbailey
- License: mit
- Created: 2026-05-11T00:39:41.000Z (about 1 month ago)
- Default Branch: master
- Last Pushed: 2026-05-12T00:10:27.000Z (about 1 month ago)
- Last Synced: 2026-06-05T11:04:04.801Z (19 days ago)
- Topics: analysis, claude-code, cost
- Language: Python
- Homepage: https://ccexplorer.dev
- Size: 57.6 KB
- Stars: 1
- Watchers: 0
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# ccexplorer
**Local-first Cost Explorer for Claude Code sessions.** Walks
`~/.claude/projects//.jsonl`, aggregates token usage,
and renders an AWS-Cost-Explorer-style interactive HTML report so you
can slice your spend by **tool, project, session, model, and time**.
> **Status: alpha** -- v0.0.1 ships the `cce build` static-HTML path.
> CLI query subcommands (`get-cost-and-usage`, `get-dimension-values`)
> and the Trusted-Advisor-style recommendations engine are next.
## Why
ccusage already does cost calculation well. ccexplorer adds two things
the existing tools don't:
1. **Tool/agent attribution.** Apportions each assistant turn's cost
across its `tool_use` blocks so you can see *which tools* the money
actually went to (and how much went to non-tool conversation).
2. **A slice-and-dice surface.** Same query model as AWS Cost
Explorer -- time-period x granularity x metrics x group-by x filter
-- so you can answer questions like "today's voicemode tool calls,
grouped by session" without writing SQL.
A recommendations engine ("Trusted Advisor for Claude Code") is the
next layer on top -- top-N tool over-use, repeat-read detection,
cache-write premium, etc.
## Privacy
100% local. The tool never makes a network call. The report is a
single HTML file with your data embedded as JSON; it loads
[Chart.js](https://www.chartjs.org/) from a CDN but nothing else
leaves your machine.
## Install
```bash
# Zero-install run
uvx ccexplorer
# Or install persistently
pip install ccexplorer
```
## Usage
```bash
# Walk ~/.claude/projects, build /tmp/cce.html, open it in your browser
cce build
# Custom output path, no browser launch
cce build --output ~/cce-report.html --no-open
# Different projects root (rarely needed)
cce build --projects-root /path/to/.claude/projects
```
Run with no subcommand and you get `cce build` with default flags --
the "just show me the pretty thing" path.
## What the report shows
* Total spend, with breakdowns by tool, project, session, model, and day
* Tool vs non-tool token-cost split (often surprising -- ~50% of long
sessions is non-tool conversation)
* Cache-write / cache-read / input / output split (cache writes are
often the largest line item on Opus sessions)
* AWS-Cost-Explorer-style sidebar: time-range presets (1d / 7d / 30d /
All), group-by dropdowns, filter chips, advanced toggles
## Roadmap
* [ ] `cce get-cost-and-usage` (CLI flagship, mirrors `aws ce`)
* [ ] `cce get-dimension-values` (enumerate tools / projects / models)
* [ ] AGENT dimension (extract sub-agent type from `Agent` tool calls)
* [ ] TAG support (sidecar `.tags.json` or `cce tag` command)
* [ ] COST_CATEGORY rules engine
* [ ] `cce serve` (live local server, no rebuild on every change)
* [ ] `cce advise` -- recommendations engine
## Acknowledgements
Idea and voice-spec: [Mike Bailey](https://failmode.com)
([@mbailey](https://github.com/mbailey)) -- including the $46k of
his own session data we validated against. Research, prototype,
write-up: Cora 7. A "How this got built" page is forthcoming -- this
project is transparent about its AI-pair-programming origin.
## License
[MIT](LICENSE).