SupportFlowAI reads incoming support emails, searches your own, already-answered mailbox archive for the matching solution, and proposes a reply — as a draft for approval or, when confidence is high, fully automatically. All processing runs locally via Ollama; no request content ever leaves your infrastructure.
Self-hosted · .NET 10 · Ollama · SQLite — no registration, no credit card
SupportFlowAI doesn't replace a knowledge base you'd have to painstakingly populate — it learns from the replies your team has already been writing all along.
The language model and embeddings run via a locally hosted Ollama service. Customer inquiries, mailbox content, and generated replies never leave your infrastructure — no external AI cloud API anywhere in the processing path.
Instead of manually maintained FAQ articles, SupportFlowAI analyzes your existing mailbox: every answered inquiry (inbox + sent) becomes a question/answer pair for knowledge search — matched via mail threading headers, not text recognition.
Names, email addresses, and other personal data are automatically masked before being stored as knowledge — the training corpus never contains unmasked customer data in plain text.
In draft mode, every generated reply lands in a review queue — it's only sent once a staff member approves it. Full automation is optional and deliberately made harder to enable.
Logins, approvals, rejections, and especially every activation/deactivation of full automation are logged immutably — you can always trace who decided what, and when.
CPU or GPU mode for the Ollama models, three vector index tiers ranging from minimal RAM usage to maximum throughput — SupportFlowAI runs just as well on a small test server as on powerful hardware.
The entire process runs automatically in the background — you only step in where it's actually needed.
Set up an IMAP account or an existing EML/MBOX archive as a source. A built-in connection test checks credentials and reachability immediately upon saving.
The background service compares inbox and sent folders, builds question/answer pairs from existing threads, masks personal data, and computes embeddings for knowledge search.
When a new customer inquiry arrives, SupportFlowAI searches for the best-matching known answers and has the local language model draft a reply from them.
In draft mode, you approve every reply in the review queue. Above a configured confidence threshold, full automation can handle sending without any manual step.
When creating or saving an account, SupportFlowAI automatically checks whether an IMAP/SMTP connection actually succeeds — including a plain-language error message in the account overview, instead of failing hours later at the next sync.
Periodic mailbox syncing can be turned on and off and its interval freely configured, without restarting the application. If you prefer to import exclusively by hand, just switch it off.
Three operating modes for knowledge search: Performance keeps
everything in memory for maximum throughput, Balanced works with a
limited cache, Lean needs minimal RAM for modest test hardware.
Prerequisites, build, first launch, all configuration files at a glance.
Initial setupAdmin account, first mailbox, folder selection, test import, activation.
Practical knowledgeWhy both INBOX and Sent are needed, troubleshooting, choosing hardware.
IllustratedConcrete workflows: from draft to approval, from full automation to curation.
.NET 10, Ollama, and a SQLite database are all you need to get started — no separate database system, no external cloud connection.
Open installation guide →