Frankie runs your support inbox.
Frankie reads every message you forward to her, labels it, drafts replies for the ones that need a quick response, files bug reports straight into your Issues queue with reproduction steps, and archives the noise. She has no ability to send: every reply she writes waits for you. The inbox stops being a pile and starts being a triaged queue with drafts ready.
What Frankie does.
Most solo founders run customer support out of personal Gmail. Customer email lands between newsletters, dentist appointment reminders, and occasional family. The customer who emailed yesterday gets buried under three Stripe receipts and a Substack digest. Some emails get answered the day they arrive; some get answered on Sunday at 11pm with apology language; some never get answered at all.
Frankie has an address of her own in your workspace. Add it as a forwarding rule wherever support mail lands today (`[email protected]`, `hello@`, whatever) and every forwarded message reaches her: no OAuth, no handing over a mailbox. She labels each one, drafts a reply for the ones that need one, files bug reports into your Issues queue with the reproduction steps drawn out of the customer's message, and quietly archives the auto-replies and the obvious spam so they never reach your morning. Approved replies go out from PorchOps's own mail domain with your business name on them, and each one carries a reply-to that routes the customer's answer straight back to Frankie.
The boundary is your forwarding rule. Frankie sees the mail you forward and nothing else, so forwarding `support@` means your personal inbox was never in reach to begin with. If you already connected Gmail before we parked that connector, it keeps working, with one caveat worth knowing: Google grants access at the account level, not the mailbox, so a connected account is the whole account. Forward from a dedicated support address rather than the inbox your dentist emails.
The drafts read like you because they're trained on your voice samples. Every labelling decision is logged with its rationale, so a wrong call is one you can find and correct. Frankie learns from the corrections, and the ones from the first week are the ones that stick.
What sets it off, what it reads, what it writes.
- Triggers
- Email forwarded to Frankie's address from your support inbox
- Inbound mail on a Gmail account you connected before the connector was parked
- A sample run, to see how she'd handle a message before trusting her with real ones
- Inputs
- Email content (sender, subject, body)
- Customer-graph match (is this email from a known customer?)
- Prior support history with this sender
- Founder voice samples
- Outputs
- A labelled message, with the reason it was labelled that way
- Drafted reply, held for your approval
- Filed Issue with reproduction steps (for bug reports)
- A quiet archive for auto-replies and suspected spam
- Audit trail entry
The playbook in plain English.
Each inbound message gets read once by Frankie. She matches the sender against the customer graph and labels the message for what it is. The labels are your workspace's own, not a fixed list we picked, so a product with a peculiar kind of inbound can name it.
Three things can then happen. A message that's clearly an auto-reply, internal, or spam gets archived without troubling you. A bug report gets filed into your Issues queue — Frankie pulls reproduction steps from the message, attaches context, and sets a sensible priority — and then still gets a drafted acknowledgement naming the Issue. Everything else gets a drafted reply. Frankie has no send permission at all, so every one of those drafts waits in the approval queue: that isn't a setting turned off, it's a capability the playbook was never given.
You see a triaged inbox: real customer questions at the top with drafted replies, bug reports already filed as Issues, vendor noise filed away. Most mornings you review the drafts, make a few edits, and send. The inbox becomes a 15-minute task instead of a constant background drain.
Every classification is reversible. If Frankie misclassifies a message — calls a real customer a vendor pitch, for instance — you recategorize and Frankie learns. The corrections themselves are the training signal.
The guardrails are yours.
- Turn Frankie off — the per-agent switch is real and enforced. A disabled agent refuses at dispatch.
- What Frankie may touch — the permissions on the installation are read on every run. Note what is absent by design: she has no send permission at all, so no setting can make her send.
- Auto-send threshold (designed, not yet wired) — would auto-send short or routine replies. Every reply holds today.
- Voice slant (designed, not yet wired) — friendly, professional, or concise.
- Escalation floor (designed, not yet wired) — the severity at or above which a message goes to you rather than getting a draft.
- Copy me in (designed, not yet wired) — the severity at which you get copied on the thread regardless.
- Classification overrides (designed, not yet wired) — domains whose mail is always labelled a particular way.
- Confidence floor (designed, not yet wired) — how sure Frankie has to be before she acts on a label instead of asking.
- Auto-archive age (designed, not yet wired) — how long untouched mail sits before it's filed away.
Frankie on support.
Frankie is the front door of the Support area. She reads every inbound message, drafts replies for the ones that need one, and files bug reports directly into Issues with reproduction steps. The acknowledgement that names the filed Issue is hers too — the bug report and the reply to the customer who sent it come out of the same run.
Read the support pillar →Common questions about Frankie.
Does Frankie touch my personal email?
Only what you forward to her. Frankie reads mail that arrives at her own address, so the forwarding rule on [email protected] is the whole boundary: nothing else in your inbox is in reach. If you connected Gmail before we parked that connector, the boundary is the account instead, since Google's permission covers the whole mailbox. Either way, use a dedicated support address.
How accurate is the classification?
We don't have a number, and we're not inventing one. Nobody has run Frankie on a real support inbox long enough to measure it. The shape we can describe: the first week is the rough one, while Frankie learns your senders and your voice, and every correction is logged and feeds the next classification. When there is a measured rate it will be dated and say whose inbox it came from.
How does Frankie handle bug reports?
When a message looks like a bug report — words like 'broken', 'not working', error messages, screenshots — Frankie pulls reproduction steps from the message body, files the Issue directly into your Issues queue with a sensible priority, and drafts an acknowledgement reply for the customer that names the Issue. You approve the reply; the Issue is already filed.
What gets escalated to me directly?
All of it, today: every draft Frankie writes waits in your approval queue, so nothing reaches a customer without passing you, and there's no separate escalation path yet. An escalation floor by severity, a copy-me-in setting, and a confidence floor are designed but not wired, and there's no keyword list either: escalating every message that says "refund" isn't something you can configure today.
Does Frankie handle non-English email?
She'll read it and draft a reply to it, in the language the model judges appropriate. There's no language detection step and no special routing, so a Spanish message doesn't get held back or flagged for you — it gets a draft like everything else. Since nothing sends without your approval, the practical answer is to read those drafts more carefully than the English ones.
What about attachments?
They're stored and she can see that a message has them, which affects how she labels it — a screenshot on a message reading "this is broken" is part of what makes it a bug report. She doesn't read inside them, so a bug described only in a PDF won't reach the Issue she files. Say it in the message body and it will.
Will Frankie archive legitimate cold outreach?
Possibly. She archives what reads as auto-reply, internal, or spam, and there's no aggressiveness dial to soften that today. A per-domain override list is designed but not wired, so for now the fix is the correction itself: recategorize the message and Frankie learns from it. Everything archived is still there and still searchable; archived is filed, not deleted.
Closing
Frankie sorts the room before I walk in. — A cofounder
PorchOps puts a crew of AI coworkers on your back office: recovering failed payments, sorting support, keeping the books. Everything they write waits for your approval.