Tone meter (0–100)
The current temperature of the thread, its trend — escalating, stable, or cooling — and the signals that drove the score. A colour spectrum so you feel the mood before you read the number.
Heated email thread? Inbox Referee reads the conversation, scores its tone, and tells you whether to reply, wait, or skip — in plain English, before you send. It runs in your browser with your own AI key, and there’s no server of ours in the middle to receive your email.
No account · No backend · No telemetry · Open source
Tone analysis
Quick replies
What it does
The current temperature of the thread, its trend — escalating, stable, or cooling — and the signals that drove the score. A colour spectrum so you feel the mood before you read the number.
One line, with the reasoning. Answers “what should I do here?” without making you scroll past six cards.
Next-action owner, so threads that are quietly stuck on you don’t slip.
Three short drafts in cooler, firmer, and warmer tones — one click to copy into Gmail.
Paste your draft, pick a target tone, get a version that keeps your intent. Stops you sending the email you’d regret.
Stale threads, “hold for me”, and archival candidates — surfaced without an extra AI call.
Thread count, escalation history, and average tone for the people you email most.
Cross-thread stats: top senders, average tone, and your “needs follow-up” list.
Your inbox is the most personal data you have. Most “AI email” tools want you to upload it to their servers. We don’t. Inbox Referee runs entirely in your browser and sends only what the model needs — the cleaned message text plus the subject, the sender and your own email address, any draft, and the message count — to the AI provider you choose, using a key only you hold. Never your contacts, your inbox, or your labels.
We can’t read your email — there’s no server of ours it ever passes through. Per OpenAI’s and Anthropic’s API terms, your inputs aren’t used to train their models.
MiniMax
Fast, cheap, recommended starting point.
OpenAI
GPT-5, GPT-4o, GPT-4.1 family.
Anthropic
Claude Sonnet, Haiku, Opus family.
Pick a model per thread. Switch any time. No lock-in.
How it works
Click the toolbar icon, pick a provider, and paste an API key. It’s validated before it’s saved.
The side panel opens with the verdict, the tone meter, and the analysis — automatically.
Click “Suggest replies” for three drafts, or drop your own draft into “Tone rewrite” first.
Honest by design
Two Chrome permissions, plus host access limited to Gmail and the AI providers you choose — nothing wider.
storage
Save your settings, your API key, the on-device analysis cache, and per-sender stats — all locally.
sidePanel
Render the analysis panel in Chrome's side panel next to Gmail.
host: mail.google.com/*
Run the content script that reads the visible thread text from the Gmail page. Nothing else on the page is accessed.
host: api.openai.com / api.anthropic.com / api.minimax.io
Make the LLM API call to the one provider you configured, using your own key, to produce the analysis and drafts.
It does not request tabs, activeTab, scripting, cookies, history, bookmarks, webRequest, all-URLs access, or any Gmail sign-in scope. The source is open — audit it.
Install Inbox Referee and let it call the tone before you do.
Add to Chrome — free