All apps
AI-native

Agentdesk

Branded AI chatbot with RAG over your docs and a clean human handoff

Agentdesk is the answer when a customer asks for "the ChatGPT, but trained on our help center, our product docs, our policies — and we want to put it on our site under our brand by next month." It is a multi-tenant chatbot with retrieval-augmented generation over a customer-private corpus, a deflection-rate analytics dashboard the support team will actually look at, and a clean handoff to a human when the bot is not enough.

Key features

  • Drop-in chat widget themed to match your site (no iframe, no third-party banner)
  • Document ingestion from URL, PDF, Notion, Google Drive, and HelpScout
  • Retrieval-augmented generation over customer-private docs (pgvector)
  • Multi-LLM gateway (Claude default, OpenAI fallback, swappable per tenant)
  • Conversation analytics with deflection rate, CSAT, and topic clustering
  • Human handoff into your existing support tool (Helpbridge, Zendesk, Intercom)
  • Per-message cost tracking with per-tenant billing rollups
  • Response guardrails and a never-answer list configurable per knowledge base

Architecture

Architecture variant: ai-native
Frontend
Next.js 14 (App Router, RSC)Embeddable chat widget (vanilla JS, no React on host page)Tailwind + headless primitives
Backend
tRPC + Next.js Route HandlersPostgres 16 with pgvectorLLM gateway (Claude + OpenAI)
Data + infra
VercelNeon PostgresCloudflare R2 for document storage
Integrations
Stripe BillingClerk for auth and multi-tenant orgsSlack and Microsoft TeamsHelpScout, Zendesk, Intercom (human handoff)

Built on the AI-Native SaaS Stack.

What you get

Agentdesk launches as a multi-tenant SaaS under your brand. Customers paste a single script tag onto their site, complete document ingestion in their first session, and start deflecting tickets the same day. The widget is themeable end-to-end — no third-party logo, no “powered by” banner.

Sample customer story (placeholder — illustrative only)

A 90-person edtech company licensed Agentdesk to handle the long tail of “how do I reset my password / where do I download my certificate / why did my charge appear twice” questions that took up 60% of their support inbox. After two weeks of live operation, deflection rate stabilized at 47% and the support team’s median response time on the remaining 53% dropped from 14 hours to under 2 — same headcount, focused on harder questions.

What’s NOT included

  • Outbound proactive messaging. Agentdesk is reactive — visitors initiate the conversation. If you need rule-triggered proactive prompts (Intercom-style), that’s a paid customization.
  • Voice (TTS / STT). Text and chat only. Voice is on the AppLiaison roadmap but not in the current box.
  • Customer-trained models. RAG over customer-private docs is in. Fine-tuning a model on a customer’s data is a paid engagement that runs on a different timeline.
  • On-prem inference. Agentdesk runs against the LLM gateway, which calls hosted Claude / OpenAI. Fully on-prem is doable but is a license-only, paid-engagement path.
  • A/B testing of agent personas. A single tone/persona configuration ships per tenant. Multi-variant testing of personas is on the roadmap.

Whitelabel surfaces

SurfaceThemeableCustom domain
Chat widget on your siteyesn/a (script-tag drop-in)
Admin web appyesyes
Email digestsyesyes (sender domain)
Webhooksyesyes
Slack and Teams appyes (logo + name)n/a

Related apps