Grounded AI chatbots, on a wholly European pipeline.
Oshu turns your docs, site, or course material into a chatbot that answers from your content. Hosted in Germany, inference in France. Nowhere else.
Where the words go.
Every message follows one short path and never leaves the EU. Two regions, no US sub-processors.
Sovereign. Grounded. Auditable.
Sovereign
A fully European pipeline, with the paperwork already done, so adopting Oshu is not a six-week procurement loop.
- EU-only infrastructureEvery hop stays in the EU. See the request path above.
- No training on your conversationsMistral's sub-processor agreement binds them; every downstream contract matches.
- Compliance, readyDPA, sub-processor list and TOMs in place from signup.
Grounded
Answers come from your library, not the model's imagination.
- Retrieval from your libraryAnswers are pulled from your own documents, not open-ended guesswork.
- Library management built inUpload PDFs, sync a GitHub repo, or crawl a web page inside Oshu. No separate backend to operate.
- A bot per siteSeparate libraries and branding for each widget you embed.
Auditable
Every action the bot takes is logged and searchable, so you can audit it rather than take it on trust.
- Per-agent analyticsConversations, resolution, latency and token use per bot.
- Conversation historySearchable, with pseudonymous session IDs.
- Full audit log + 12-month auto-purgeEvery action recorded. Purged after 12 months per the retention policy.
A grounded answer, and the operator’s view.
What your visitors see, and what you see. Both in one place.
The two views above are an illustration. The chat launcher in the bottom-right corner is the real Oshu widget, the same embed you would ship, answering from this site’s own docs. Open it and ask anything.
Everything the agent can do.
GitHub Markdown sync
Point an agent at a repo and folder; new commits are pulled into its library automatically, on a schedule.
PDF upload with OCR
Drop in PDFs and we OCR them into Markdown for the agent's library. Scanned, image-only documents work too.
Web page sync
Point a library at a web page or site section; we crawl it, extract the text, and re-index it on a schedule.
LaTeX & math rendering
Inline and block KaTeX renders equations, derivations, and scientific notation correctly inside the chat replies.
Grounded answers
Answers come from your library, not the open web; reveal each source to visitors, or keep them dashboard-only.
Visible reasoning
Optional: Premium and reasoning models think step by step, and the widget shows a live, collapsible thinking view.
Content moderation
Optional guardrails screen visitor messages by category, then warn, block, or end the conversation per your rules.
Streaming replies
Replies stream token by token over SSE, so visitors watch the answer take shape live instead of waiting for it.
Theming
Brand colours, fonts, radius, and labels are configurable per agent, so the widget looks like part of your site.
Embed layouts
One script tag drops the agent in. Pick a floating button, sidebar, inline panel, full page, or iframe per embed.
Hosted page + QR
Turn any agent into a standalone page at its own link, and share it with the QR code built into the dashboard.
27 languages
Replies match the question's language; the widget UI auto-detects the visitor's browser locale in 27 languages.
Private mode
Switch an agent to private and visitor messages aren't stored; each turn logs a placeholder so usage stays exact.
Conversation analysis
Optional: finished chats are reviewed for helpfulness and flagged for knowledge gaps, so you know what to add next.
Operator dashboard
Every conversation, message volume, latency, model split, and a per-agent audit log, all in one workspace.
Five ways to embed.
One script tag, your pick of layout. The same agent drops in as a corner bubble, a side panel, a full page, an inline element, or an iframe. Or skip the embed entirely: enable a hosted Oshu page for the agent, with its own link and QR code to share.
Floating
Chat bubble in the corner.
Sidebar
Slides in from the side.
Full page
Takes over the page.
Inline
Inside a page element.
Iframe
Embedded hosted page, great for an LMS.
Flat monthly fee. Hard cap.
One flat price a month, no per-message meter. Caps are hard: at the limit the bot pauses instead of running up a surprise bill. And if you need more headroom, we can raise any cap.
Caps are counted in message credits. On every plan the base model stays 1 message credit per reply; Premium and any future models may use a few more.
- Embed widget, full theming
- Base & Premium model tiers
- No card required
- One agent, full theming
- Grounded, LaTeX, web + GitHub + PDF
- Hard cap, optional extension
- Three agents, three sites
- Grounded, LaTeX, web + GitHub + PDF
- Hard cap, optional extension
- Up to ten agents
- For docs, support, onboarding
- Hard cap, optional extension
- 50 agents per workspace
- Agencies & multi-product teams
- Hard cap, optional extension
- Bring your own Mistral key
- Unlimited messages + library
- Per-agent model picker
For organisations that need more than a plan: several users per workspace with roles, tailored caps and SLAs, and managed onboarding, where we set up the agents, libraries, and widget for you. Self-hosted or white-label available.
Cheaper than US-hosted alternatives, and this one keeps your data in the EU.
Prices are net, excl. VAT. 19 % German VAT is added at checkout for German customers. EU businesses outside Germany with a valid VAT-ID: reverse-charge (no VAT). Non-EU: no VAT.
Now live
Put a grounded, European bot
on your site this week.
Flat price, hard cap, DPA already in place. Start free, no card, no per-message meter.