AI agents that close the loop

Most AI integrations stop at "draft a reply, then a human sends it." Helploom ships two Model Context Protocol servers that go further — your AI reads, drafts, sends, and oversees, on its own. Bring your own model. No LLM markup.

Two MCPs, one workflow

Support Agent MCP — the agent. Your AI replies to customer tickets in real time, with your voice, against your knowledge base. Twelve tools including reply_to_chat, add_internal_note, and engineering-routing for product bugs.

Support Manager MCP/B — the overseer. Your AI runs a daily review of your queue: aging tickets, agent performance, escalation routing, voice-of-customer themes. Works with your existing helpdesk — Zendesk, Freshdesk, Intercom, Helpscout, or Helploom. Free, bring your own AI, install in Claude Desktop in under a minute.

Support Agent MCP — for orgs on Helploom

Connect Claude, GPT, Gemini, or your own agent stack to your Helploom org. Your AI doesn't just read — it acts.

What it does

• Reads every conversation — full thread, customer history, RAG-retrieved KB articles, similar resolved tickets as gold-standard examples.
• Drafts replies in your voice — your voice guide, signature, and few-shot examples ride inside every context call. Drafts sound like your team wrote them, across Claude, GPT, and Gemini.
• Sends replies, with your permission — reply_to_chat mirrors out via the native channel (widget, email, Slack). Capability flags and per-token rate limits keep the blast radius bounded.
• Routes product bugs to engineering — flag_chat_as_engineering plus create_or_link_engineering_ticket push real defects to Linear, Jira, or GitHub Issues with de-duplication.

Why it's different

• Closes the loop. Intercom's MCP and Pylon's MCP let agents read everything but can't send customer-facing replies. Ours does.
• Routes bugs separately. Most "support AI" tooling drops bug reports back into the support queue. Ours splits the inbox: support questions go to support, defects go to engineering.
• No LLM markup. You're already paying Anthropic, OpenAI, or Google. We don't tax you a second time for inference.

Support Manager MCP/B — for any helpdesk

You don't have to be on Helploom to use the Support Manager. Install the bundle in Claude Desktop, point it at your existing helpdesk, and your AI runs the oversight layer your team doesn't have time for.

What it does

• Daily queue digest. Unresolved aging, response-time trends, top tags, handoff counts — composed by your AI, delivered to whoever you choose.
• Reply quality scoring. Five-criteria rubric (accuracy, empathy, tone, grammar, personalization) applied to your team's replies. Per-agent coaching summaries, weekly.
• Escalation routing. Real bugs go to engineering, billing issues go to finance, the long tail goes to the right specialist — by email, auto-tag, and Slack, not phone calls.
• Voice-of-customer themes. Your AI rolls up recurring complaints into a weekly report you can hand to product.

Works with your helpdesk

Helploom, Zendesk, Freshdesk, Intercom, Helpscout. One install, one helpdesk, uniform tool surface regardless of backend.

Why it's free

The Support Manager MCP/B doesn't cost anything because we're not running the inference. You bring your AI client (Claude Desktop, ChatGPT, Cursor, agent SDK) and your helpdesk credentials. Helploom is the protocol layer; your tokens pay your model provider directly.

If you eventually want this as a managed service — daily digests arriving in your inbox without you having to ask Claude every morning, escalations firing automatically on schedule — that's our hosted Support Manager, priced by usage. For now, the bundle is the way in.

What we capture (and what we don't)

We capture: which helpdesk you connected (Zendesk, Freshdesk, Intercom, Helpscout, or Helploom). That's it — operational, so we know which integrations to support.

We never see: your ticket content, your customers' PII, your agents' replies, your internal notes, your tool-call counts, or anything that flows through your AI client. The QA scoring runs locally; we don't see the prompts or completions.

Why this matters

Support teams have been told "AI will replace front-line agents." That's not what's happening. AI replaces the parts of support that are scripted — investigation playbooks, draft replies, queue triage, quality QA, escalation routing — and frees humans to do the parts that require judgment.

Our MCPs make that division of labor real. The Support Agent handles the per-ticket work. The Support Manager handles the manager-layer work — the stuff your head of support is doing by hand today. Together, they let a 3-person team operate like a 15-person team.

And because you bring your own model, you're not paying us for inference you're already paying Anthropic, OpenAI, or Google for.

FAQ

Do I have to use Helploom as my helpdesk? No. The Support Manager MCP/B works with Zendesk, Freshdesk, Intercom, Helpscout, or Helploom. The Support Agent MCP requires a Helploom org because it operates on Helploom's data model.

Which AI clients work? Anything that speaks MCP. Claude Desktop, Claude Web, ChatGPT (with MCP support), Cursor, agent SDKs from Anthropic, OpenAI, or Vercel, LangChain — anything in the MCP ecosystem.

Is it really free? The Support Manager MCP/B is free. You pay your AI provider directly; we don't take a cut. The Support Agent MCP is included with your Helploom subscription — no surcharge.

What about EU data residency? US-first. EU and AU regional hosting follow customer demand.

How is this different from Intercom's MCP or Pylon's MCP? Theirs are read-only and propose-only. Ours sends. Theirs don't separate bugs from support; ours route bugs to your engineering tracker. Neither has a manager-layer product.