Agent online

Support and sales forhumans & AI agents.

Your customers ask ChatGPT, Claude, and Perplexity now — and those agents shop your site on their behalf. HelpShelf is one autonomous agent that supports and sells to both: cited answers in your widget, inbox, and help center, plus a live MCP + llms.txt surface so AI agents get it right too. It taps you only when it's stuck.

Start free — no card
Supports + sells
one agent
Every surface
humans + AI agents
Hands off
only when it’s stuck
helpshelf · agent demoresolved
Scenario 1/3Support · website visitor

▸"Do you support SSO on the Pro plan?"

⚙search_docs("sso pro plan") → 3 cited sources

Yes — SAML SSO is on Pro and Scale. [1][2]

✓confidence 0.94 → auto-sent, no human needed

support · sales · ai agents

One agent, three jobs — support, sales, and the AI agents shopping for your customers.

For developers

Live in one command

Run it in your terminal and HelpShelf scans your codebase — routes, docs, components, tests — then ships an agent that supports and sells, answering with citations. Works with npx and bunx, on any stack.

npx helpshelf init
npx helpshelf scan
npx helpshelf push
Start free — no cardSee the CLI
Terminal
npx helpshelf init
✔ Created .helpshelf/
npx helpshelf scan
✔ Next.js · 14 routes · 28 deps · 180 UI strings
✔ FAQs: 6 generated
npx helpshelf push
✅ Agent online — answering with citations
01Surfaces

Everywhere your customers actually ask.

One agent, one knowledge base, one answer — delivered in your widget, your support inbox, your help center, and inside the AI assistants your customers already use. Human or AI, everyone gets the same cited answer.

  1. 01In-product widget
  2. 02Support inbox
  3. 03Help center
  4. 04ChatGPT · Claude · Perplexity
Floatie · acme.com

Do you support SSO on the Pro plan?

Yes — SAML SSO is included on Pro and Scale. Setup takes about 10 minutes.[1][2]
Re: SSO question — support@acme.com

Hi Dana — yes, SAML SSO is included on Pro and Scale. Setup takes about 10 minutes; guide below.[1][2]

drafted by your agent · approved from Telegram

help.acme.com/sso-setup

Setting up SAML SSO

SSO is available on Pro and Scale plans. Connect your identity provider in Settings → Security…

kept fresh by nightly crawl + gap detection

ChatGPT — "does acme support sso?"

Acme supports SAML SSO on Pro and Scale plans, with a ~10 minute setup.

source: acme's llms.txt + live MCP endpoint — served by HelpShelf

02Receipts

Every answer shows its work.

Black-box AI loses customer trust the first time it's wrong. HelpShelf attaches inline citations to every answer and ranks sources by trust tier — so you, your team, and your customers can check the evidence.

answer · assembled from your sourcesverified

Yes — SAML SSO is included on the Pro and Scale plans.1

You can connect Okta, Azure AD, or any SAML 2.0 provider from Settings → Security.2

Most teams finish setup in about 10 minutes.3

[1] Security → Single sign-oncurated
[2] Settings reference — identity providerscurated
[3] Onboarding guide (crawled)standard

Trust tiers — how sources are ranked

Curated 1.3×

Hand-written or imported from your help desk. Human-reviewed, ranked first.

Standard 1.0×

Crawled from your site and docs. The reliable baseline.

Generated 0.8×

AI-drafted, pending review. Deprioritized until a human approves it.

Hybrid search (BM25 + vector, fused by reciprocal rank) applies these multipliers on every query — so human-reviewed answers always outrank AI drafts.

Telegram · HelpShelf bot · 11:04 AM

Your agent needs you

Dana @ Acme — refund for a March order

“I couldn't resolve this one — it needs account access and refunds are on your always-hold list. Full conversation and retrieved context attached.”

31 resolved today·this is the 1st that needs a human

Why it reached out — exactly

  • · needs account access (can't be answered from docs)
  • · topic matched always-hold list: “refunds”
  • · everything else today: handled automatically
03Escalation

It only needs you when it doesn't know.

Confident, well-sourced answers send themselves. The agent pings your Telegram, Slack, or email only when something genuinely needs a human — with the whole conversation and its retrieved evidence attached, so you start from context, not from scratch.

  • Autonomous by default. Replies that clear the quality bar — high confidence, multiple sources, human-reviewed content — go out automatically.

  • Reaches out only when stuck. Low confidence, weak sources, or account-specific requests escalate to you with full context.

  • Always-hold topics. Refunds, security, billing exceptions — name the topics a human should always handle, and the agent never freelances them.

  • Review-everything mode. Want to approve every reply while you build trust? Flip one setting and each draft waits for your one-tap OK.

04Sales

It notices when support turns into sales.

Half of “support” questions are buying questions in disguise. HelpShelf classifies intent on every message — when a visitor asks about pricing, plans, or how you compare, the same agent switches into sales mode and answers like your best rep: grounded, cited, and pointed at the next step.

  • Automatic intent detection. Every message is classified sales or support — no setup, no separate bot to configure.

  • Pin sales mode where it counts. Page rules lock the agent into sales mode on your pricing and comparison pages.

  • Answers that close, with receipts. Plan comparisons and feature questions come back cited from your real docs — not hallucinated promises.

  • Sales readiness gauge. Your dashboard scores how prepared the agent is to sell, with the next action to improve it.

floatie · acme.com/pricingmode: sales

How do you compare to Intercom? We're a team of 6.

For a 6-person team, the practical difference is pricing model and setup time: we charge per resolved conversation instead of per seat, and most teams are answering from their own docs in about 10 minutes.[1][2]Want me to spin up a free agent on your docs so you can compare answers side by side?

intent: sales (0.91) · mode pinned by page rule · sources: pricing [1], comparison guide [2]

05The inbox

It works the inbox so you don't.

Connect Gmail and HelpShelf runs real support email end to end: it classifies every message, sends confident cited replies on its own, flags what it can't answer as knowledge gaps, and sums up the day in one digest. Quiet days stay quiet.

92/100

Answer readiness

90/100

Sales readiness

88/100

Agent readiness

Readiness gauges live on your agent's home — with the next best action to raise each one.

your agent's day · while you shipped product

  1. 02:00

    Nightly crawl — 4 docs updated, 1 knowledge gap detected

  2. 06:12

    Replied to Dana @ Acme — auto-sent, confidence 0.94, 2 curated sources

  3. 09:30

    14 widget questions resolved with citations — 3 were sales intent

  4. 11:04

    Pinged you once: a refund request that needs account access

  5. 13:00

    Daily digest: 31 resolved on its own, 1 handoff, 1 gap to fill

06AI customers

Your next customer might not be human.

People ask ChatGPT, Claude, and Perplexity about your product — and increasingly, AI agents research and recommend purchases on their behalf. HelpShelf makes you legible to them: a spec-correct llms.txt plus a live MCP endpoint, so AI agents get your verified answers, your pricing, and a real path to your team.

  • llms.txt + llms-full.txt generated and kept fresh automatically
  • Live MCP server — AI agents search your docs, ask questions, and pull verified product context
  • AI agents can escalate to you: escalate_to_human creates a real handoff, straight to your Telegram
  • Curated and standard sources only — AI drafts never leak out
Generate your llms.txt — freeSupport ROI calculator
acme.com/llms.txt200 OK · served by HelpShelf
# Acme Rockets> Model rocket kits, telemetry, and launch support.## AI Agent Access- [llms-full.txt](https://help.acme.com/llms-full.txt)- MCP: app.helpshelf.com/api/mcp/f364256c  tools: search_docs, ask, get_article,  get_product_context, escalate_to_human## Docs- [Assembly guide](https://help.acme.com/assembly)- [Telemetry setup](https://help.acme.com/telemetry)- [Pricing & plans](https://acme.com/pricing)
07Builders

You vibe-coded the product. Don't hand-code the support.

Built with Lovable, Replit, Cursor, Claude Code, Bolt, or v0? You never left your AI builder to ship the app — don't leave it to staff the app. Grab your free site key, paste one prompt, and your product has an agent that supports and sells.

  • One script tag — your AI builder places it for you
  • HelpShelf crawls the site you just shipped and builds its own knowledge base
  • Connect GitHub or run the CLI to teach it your codebase too
HelpShelf for vibe coders

The one-paste install

Paste this into the Lovable chat:

Add HelpShelf to this app: load
https://app.helpshelf.com/widget/MY_SITE_HASH.js
with an async <script> tag just before the closing
</body> tag on every page. Replace MY_SITE_HASH with
the site key from my HelpShelf dashboard.

Your site key takes 60 seconds: start free, point HelpShelf at your site, and it does the rest.

08Pricing

Pay for resolutions, not seats.

One credit = one conversation your agent fully resolves. Escalations, low-confidence answers, and noise are free — you only pay when the agent actually does the work.

✓ resolved → 1 credit↗ escalated → free✗ unanswered → free

Starter

Try a real agent on one product.

$0/mo
Start free
  • 1 agent
  • Cited answers + trust tiers
  • Widget + help center
  • Optional review-everything mode
Most teams choose this

Pro

For teams handing support to the agent.

$49/mo
Start Pro
  • 3 agents
  • Autonomous email inbox + Telegram/Slack handoffs
  • llms.txt + MCP AI surface
  • Custom branding

Scale

For companies running support on autopilot.

$149/mo
Start Scale
  • Unlimited agents
  • Highest resolution volume
  • Priority support
  • Advanced analytics

Full plan details on the pricing page.

09FAQ

Answers about the answer machine.

What is an AI customer service agent?
An AI customer service agent is software that resolves customer questions automatically using your product documentation, website, and codebase as its knowledge source. HelpShelf goes further than a chatbot: it retrieves evidence before answering, cites every source inline, handles sales questions as well as support, and escalates to a human only when it cannot resolve something on its own.
Does HelpShelf run on its own, or do I approve every reply?
It runs on its own. Replies that clear a strict quality bar — high confidence, multiple sources, at least one human-reviewed — send automatically. The agent only reaches out (via Telegram, Slack, or email) when it is stuck: low confidence, account-specific requests, or topics you marked always-hold like refunds. If you prefer to approve everything while you build trust, flip one setting to review-everything mode.
Can the same agent handle sales questions?
Yes. HelpShelf classifies the intent of every message as sales or support automatically. Pricing, plan-comparison, and "how do you compare to X" questions get sales-mode answers grounded in your real docs, and you can pin sales mode on pricing pages with page rules. A sales readiness gauge in the dashboard shows how prepared the agent is to sell.
How does HelpShelf keep answers accurate?
Every answer is assembled from retrieved sources ranked by trust tier — curated (human-reviewed, boosted 1.3×), standard (crawled), and generated (AI-drafted, penalized 0.8×). Answers carry inline citations, and anything below the confidence-and-sources bar is held for a human instead of being sent.
Can the agent answer my support email?
Yes. Connect Gmail and HelpShelf triages incoming support email end to end: it classifies each message, sends confident cited replies automatically, holds the rest for a one-tap decision, and rolls the day up into a digest.
What does it mean to serve AI agents as customers?
People increasingly ask ChatGPT, Claude, and Perplexity about products — and AI agents research and recommend purchases on their behalf. HelpShelf publishes llms.txt and llms-full.txt for your site and runs a live MCP endpoint, so AI agents can search your docs, ask questions, pull verified product context, and even escalate to your human team with the escalate_to_human tool.
I built my app with Lovable, Replit, or Claude Code — how do I add HelpShelf?
Without leaving your AI builder. Sign up free to get your site key, then paste one prompt into Lovable, Replit Agent, Cursor, Claude Code, Bolt, or v0 telling it to add the HelpShelf widget script tag. HelpShelf crawls the site you shipped to build its knowledge base, and you can connect GitHub or run the CLI so it learns your codebase too.
How does credit pricing work?
One credit equals one conversation the agent fully resolves on its own. Conversations that escalate to a human, get no answer, or are spam cost nothing. You pay for outcomes, not seats or message volume.

Five minutes to first cited answer

Install it. Then go build something.

Point HelpShelf at your site, docs, or repo. It builds the knowledge base, supports and sells with cited answers — to humans and AI agents — and only taps you when it needs you.

Start free — no cardView pricing