Runs on your device

How do you talk to your AI?

AI Karma reads your AI chat history back to you: how often you say please, thanks, or something sharper. It runs on your device.

Try it on this deviceNo sign-up. Your files stay in this browser.

Two ways to use it

Available now · default

On this device

You pick history files. Your browser reads and analyses them.
Conversations, results and exports never leave the device.
No account, no analytics, no AI models: the score comes from fixed rules.
Data stays in this browser until you clear it. Export an archive to keep a copy.

Later · optional

Connected

Sign in and combine histories from several devices.
Only after you agree to upload, and only the files you pick. Our server can read uploaded data to process it.
Optional public profile and optional model-based analysis, each with its own consent.

Opening this website downloads the app, and the host may keep ordinary access logs (for example your IP address). The on-device promise covers your conversations and results: those are never sent.

A sample report

Invented numbers so you can see the format. Pick a score to see how the signed scale reads, from −100 to +100, with 0 in the middle.

Demo · sample data, not yours

Demo AI Karma

AI Karma +67 on a scale from −100 to +100

The Ray of Sunshine

Thanks and praise everywhere. Your assistant would probably write you a thank-you note.

A playful score based on the words you used. It covers only the files you imported, and it isn’t a verdict on you.

What moved the needle

Your signature moves

17messages said thanks, out of 371
16requests came with a please
18corrections (fair game, not counted against you)
7messages aimed an insult at the assistant

Kind vs. harsh

Where your words landed

Plain, direct instructions count as neutral. Corrections and frustration are never counted as harsh. Readable messages: 371.

Activity

The ground covered

Your messages: 412
Conversations: 58
Active days: 41

What we read

371 of your 412 messages were readable

That’s 90% of your messages. The rest were set aside, never counted against you:

Language we can’t read yet (English only for now)12
Only code, quotes or empty29
No date in the source7
Not clearly written by you2
Assistant replies, tools and system text (never scored)2,980

Missing dates and unknown authors are disclosed, never guessed. Nothing here counts as harsh.

Method

How this was worked out

Fixed, versioned rules look for polite requests, thanks, praise, apologies and clearly targeted insults in your messages. No AI model is involved. A score appears once there are at least 30 readable messages and 10 messages with a signal.

score = round(100 × (P − H) ÷ (P + H + 10)), kept between −100 and +100

Karma processor: 1.0.0
Rules: karma-en-1.0.0
Text extraction: extract-1
Language check: lang-heuristic-1

Demo data: invented for illustration.

What you can import

Each source says how well it has been checked. “Unverified” means we built it from documentation and have not tested it against a real export yet.

Supported sources and their verification status
Source	Status	Where to find it
Claude Code	Tested on real local data	Session files in ~/.claude/projects (one folder per project).
Codex CLI	Tested on real local data	Session files in ~/.codex/sessions and ~/.codex/archived_sessions.
Claude (web export)	Unverified	The conversations.json file from a claude.ai data export, unzipped.
AI Karma archive	AI Karma format	A folder you exported from AI Karma (ZIP archives: coming soon).
Ollama desktop	Deferred	Not supported yet.