Claudon

See where your Claude Code time and money actually go.

One command turns your session logs into an interactive HTML dashboard.
100% local · zero dependencies · nothing leaves your machine.

uvx claudon --open

_{Redacted sample. Your own sessions stay on your machine.}

Quick start · In the browser · Usage · On-device AI · Privacy · Contributing

Why Claudon?

Claude Code is fast until it isn't. Which turn burned ten minutes? Was it the model, a slow tool, or you being away? Did a failing tool call loop? What did the session really cost?

The answers are already in ~/.claude. Claudon reads them and shows you:

It is a small Python package using only the standard library. No telemetry, no network, no account.

Quick start

Run it with whichever tool you already have. Each one downloads the latest release on demand.

After a global install:

claudon ~/.claude -o report.html --open

Requirements: Python 3.11+ on your PATH (python3, python or py). The npm-family packages are a thin launcher around the same Python script. No Python? Use the browser version or a standalone binary from GitHub Releases.

curl -fsSL https://raw.githubusercontent.com/infenia/claudon/main/install.sh | sh

# recommended: pin a release (see GitHub Releases); the download is checked against that release's SHA256SUMS
curl -fsSL https://raw.githubusercontent.com/infenia/claudon/main/install.sh | CLAUDON_VERSION=vX.Y.Z sh

irm https://raw.githubusercontent.com/infenia/claudon/main/install.ps1 | iex
# pin a release: $env:CLAUDON_VERSION = 'vX.Y.Z' first

git clone https://github.com/infenia/claudon.git
cd claudon
python3 -m claudon ~/.claude -o report.html --open

uvx --prerelease allow claudon
npx @infenia/claudon@beta

In the browser (no install)

Claudon also runs entirely in your browser, with no local Python.

• Live app: claudon.infenia.com/app
• Locally: from the repository root run python3 -m http.server, then open http://localhost:8000/wasm/.

Drop .jsonl transcripts onto the page, or use Open Folder on ~/.claude/projects to keep project grouping and subagents. The optional AI summary and Ask AI work here too, in browsers that support the built-in AI APIs.

Use it inside Claude Code

claudon --install-plugin

Then type /claudon in any Claude Code session to generate and open your dashboard.

Usage

claudon [PATH] [-o FILE] [--redact] [--pricing FILE] [--open] [--install-plugin [--force]] [--version]

PATH can be a ~/.claude directory, its projects/ subfolder, a specific project directory, or a single .jsonl transcript. The default is $CLAUDE_CONFIG_DIR if set, else ~/.claude.

Common recipes

claudon --open                                  # analyze everything, open the report
claudon ~/.claude/projects/my-project --open    # one project only
claudon --redact -o share-me.html               # safe to attach to an issue

What's in the report

The report is one self-contained .html file: no server, no external assets.

The Bottlenecks findings

Every finding is computed exactly from the transcripts (no AI involved) and comes with a Fix line:

• Idle time and permission-prompt waits (Read/Edit/Write calls that sat waiting for your approval)
• Failing tools, and tasks stuck retrying the same failing step
• Repeated identical reads/searches and edit churn (the same file edited 10+ times)
• Context bloat past 400k tokens, expired prompt caches after a 5+ minute pause, and a high cache-write share
• Interrupts and rejected tool calls, heavy thinking, and subagent cost share
• Tasks with unusually many turns for their project, cost concentration, and trends (earlier vs. later tasks)

Filters (project, search) apply to every finding. The thresholds live in one RULES object in claudon/dashboard.html.

On-device AI (optional)

When your browser can run its own built-in language model, the report adds:

• AI summary (Overview tab, under the KPIs): a headline, where the time and money go, and up to four priorities. Each priority opens the Bottlenecks finding it is based on.
• Ask AI (its own tab, Prompt API only): chat about the sessions in view, e.g. "Why was yesterday slow?", "Which tools fail most?" or "What did expired caches cost me?". Tasks the answer cites become clickable, and Data the model saw shows exactly what it was given. The chat is kept in your browser's local storage (per report) until you press Clear chat.

Which browsers

Support is detected by asking the browser itself (LanguageModel.availability() / Summarizer.availability()), never from its name, so new browsers work as soon as they ship the APIs.

Hardware. Chrome runs Gemini Nano on Windows 10/11, macOS 13+, Linux or ChromeOS (Chromebook Plus). It needs about 22 GB of free disk, plus either a GPU with more than 4 GB of VRAM or 16 GB of RAM and 4 CPU cores. Edge has similar requirements for its Phi model. If the browser reports that the device can't run the model, the AI features stay hidden. Check the model's state at chrome://on-device-internals (Edge: edge://on-device-internals).

First use. The browser may need to download its model once (a few GB). Claudon never starts that download on its own: it shows Enable on-device AI, and the browser downloads its model only after you click. After that the model runs locally and keeps working offline.

Where it works

• A report the CLI wrote, opened from disk (file://)
• The browser app, whose report frame is allowed to use the AI APIs
• A report saved from the browser app

How it stays accurate. The AI never computes numbers. It is given the findings and report data (already-formatted figures), is told to use only that data, and can only point at findings that exist. Its output is shown as plain text. For chat, the data for each question is picked by fixed rules (the tasks, projects, tools, dates and metric it mentions), not by the model. The numbers in the findings are always exact; treat the AI's wording as a summary.

Not seeing it? Your browser doesn't expose the APIs, the device doesn't meet the requirements, or (Edge) the Prompt API flag is off. The live check on the project page tells you which features your current browser supports.

Privacy

Everything runs on your machine. The CLI makes no network requests, and neither does the report page. The optional AI summary and Ask AI chat use the browser's own on-device model, so your data never leaves your computer for them either. If the browser still has to fetch that model, it does so itself, and only after you click Enable on-device AI. Chat history stays in your browser's local storage until you clear it.

The browser version uploads nothing either: files are read and analyzed locally. Its only downloads are the Pyodide runtime (from jsDelivr) and claudon.pyz, and once it shows Ready it keeps working offline. Chrome and Edge ask to let the page view the folder you pick. Firefox and Safari show their own "upload" prompt for folder pickers, which only grants the page read access.

Releases

Claudon follows Semantic Versioning. Conventional Commit PR titles drive an auto-generated release PR (version bump + CHANGELOG); merging it tags the release and publishes:

• PyPI (trusted publishing) and npm (with provenance)
• Standalone binaries for Linux x86_64, macOS (arm64 and Intel) and Windows, on GitHub Releases with SHA256SUMS and build provenance attestations (gh attestation verify <file> -R infenia/claudon), plus claudon.pyz (single-file Python bundle, used by install.sh where no binary exists)

Full process: docs/RELEASING.md.

Contributing

Contributions are welcome. Please read CONTRIBUTING.md (PR titles must be Conventional Commits) and follow the Code of Conduct. Report vulnerabilities privately per SECURITY.md.

License

Developed by Infenia Private Limited. MIT License. See LICENSE.