Build a chatbot

A SeldonFrame chatbot is a Soul (who it is) plus a Brain (which LLM and key) plus tools (what it can do). Build one in a single Claude Code prompt.

The fastest path

Open Claude Code and say what the chatbot is for. Be specific about the business, the persona, the tasks, and the price points. SeldonFrame's build_website_chatbot tool bundles everything — create, configure, eval, publish — into one round-trip.

> Build me a website chatbot for "Acme Dental." It should:
  - greet patients warmly, never aggressively
  - answer FAQs about cleanings ($120), fillings ($200-450), whitening ($350)
  - book appointments using my Google Calendar
  - never quote prices outside this list — say "I'll have the office confirm"
  - escalate emergencies (broken tooth, severe pain) to call (555) 010-0100

Claude Code generates the Soul (personality + system prompt + FAQ snippets), wires the booking tool, generates an 8-scenario eval suite, runs evals, and reports the pass rate. If it's ≥87.5%, you can publish.

What you can configure

Persona (Soul)

Name, role, tone, greeting, sign-off. Trusted-source allowlist (which facts the bot is allowed to repeat without checking). FAQ snippets. Pricing rules. Refusal rules ("never give medical advice").

Brain

Pick the LLM provider (Anthropic / OpenAI) and the model (Sonnet 4 / Opus 4 / GPT-4 / etc.). Uses your own API key — see Settings → LLM keys.

Tools

Pick which capabilities the bot has: book appointments, reschedule, cancel, look up customer history, send a follow-up email, etc. Each tool is one MCP call.

Eval suite

8 default scenarios cover greetings, FAQ accuracy, booking, rescheduling, refusals, PII handling, escalation, and tone consistency. You can add your own — Claude Code can write scenarios from a real conversation transcript.

Manual UI flow

Prefer clicking? Open Agents and hit "New agent." You get the same five tabs:

Overview — the agent's stats and live status.
Sandbox — chat with it in isolation.
Conversations — read what your customers said to it.
Settings — Soul, Brain, tools.
Evals — run the eval suite, gate publish.

Skill-pack architecture

Agent intelligence (greeting style, refusal patterns, escalation logic) lives in markdown skill-packs the prompt composer reads at runtime — not in hard-coded TypeScript. So when Claude or GPT improves, your agent improves too. No rewrites required.

Going live

Run evals → if pass rate ≥87.5%, the "Publish" button unlocks → the agent flips to live status. Now its embed snippet is hot. Drop it on any page.