Introducing Olmbox: an AI inbox that never sends for you
Most “AI for your contact form” tools are built around automation: a message arrives, a model writes something back, and it goes out. That is a reasonable product. It is not this one.
Olmbox analyses every Contact Form 7 submission and drafts a reply — then stops. The draft sits in an inbox until a human reads it and clicks send.
Why draft-only is the whole design
The failure mode of automated replies is not that they are bad on average. It is that the worst one gets sent too. A model that writes a good reply to 95 out of 100 enquiries will confidently mishandle the refund request, the legal threat, or the message from someone in distress — and it will do so in your company’s voice, before anyone has read it.
Reviewing 100 drafts is much less work than writing 100 replies, and it costs you nothing on the 95 that were fine.
So the constraint is architectural rather than a setting. There is exactly one code path that emails a visitor, and it runs only in response to a confirmed click in the admin. At submission time the plugin does nothing but write a row to its own table. There is no toggle to turn that into automatic sending, because the moment such a toggle exists, the guarantee stops being a guarantee.
What one API call buys
Each submission produces five things: a summary, a suggested reply, a category, a priority and a confidence score with reasoning.
These come back from a single request. The model is asked for one JSON object containing all five, rather than being asked five times. Triage that is five times richer therefore costs the same as triage that is not, which is what makes it reasonable to run on every submission rather than only the ones you already suspect are important.
Confidence is a feature, not decoration
The score matters most when it is low. A short, vague message — “hey do you do the thing for the bigger plan?” — is exactly where a drafted reply is most likely to guess wrong, and it is exactly where a confident-looking draft would be most dangerous if it sent itself.
Anything under 60% is flagged in the list. The flag is not a warning that the plugin is broken; it is the plugin telling you which three of today’s forty messages deserve your full attention.
Failures should cost you the analysis, never the message
If the provider times out, the submission is still captured, still appears in the inbox, and still tells you what went wrong. Contact Form 7’s own notification email has already gone out by then, untouched.
An AI feature that can lose a customer enquiry when an API has a bad afternoon is worse than no AI feature. The ordering here is deliberate: Olmbox runs after CF7 has finished its own work, so there is nothing it can interrupt.
Bring your own key
There is no resold access and no subscription. You supply a key for OpenAI, Anthropic, Gemini or OpenRouter, it is encrypted at rest in your own database, and no request is made to anyone until you have configured one.
No telemetry is collected. Nothing is sent to us, because there is no “us” to send it to — the plugin has no server component at all.
Where it is
The plugin is in review for the WordPress.org directory. The source is on GitHub, GPL v2-or-later, and the documentation covers installation and the review workflow.
If you try it, the thing worth reporting back is not whether the replies are good — it is whether reviewing them is faster than writing them. That is the only claim the design actually makes.