Support tools with an MCP server, compared

An MCP server lets an AI tool like Claude, ChatGPT or Cursor read your support tool’s conversations and tickets when you ask it something. Many can change things too, and some can reply to your customers. Most of the big support tools now have one. Zendesk, the best known, has announced its server but hadn’t said how to use it when we looked.

We read the MCP docs of 32 support tools on 2 November 2026, and link each fact to the page we read it on. Of the 32, 23 have a server for your support data that says what it does. Wonderchat has one it hasn’t documented, Zendesk’s is on the way, three have one built for another job and four have none.

Deacon comes first, because we make it. It’s an AI customer support agent that answers your customers from your own docs. Its MCP server is for your coding agent rather than a support team. Ask Claude Code or Cursor “what broke this week?” and it reads what your customers reported, beside the errors your site logged, then shows you the fix in your code. It’s on every plan, the free one included.

Support tools with an MCP server, side by side

The column that matters most is whether it can reply to your customers, because a sent reply can’t be taken back. What each one reads and changes is in its own section below.

What a support tool’s MCP server is for

MCP, the Model Context Protocol, is how AI tools like Claude, ChatGPT, Cursor and Claude Code connect to other software. Two kinds of people use a support tool’s server. A support team asks Claude or ChatGPT to sum up this week’s tickets or draft a help article. A developer asks a coding agent what customers reported, and has it look at the code. Most servers here are built for the first, which is why so many sit in Claude’s and ChatGPT’s own directories.

Some of these tools also work the other way round. Zendesk, Intercom, Freshdesk, Front, HubSpot, Missive, Crisp, Tidio, Plain, Pylon and tawk.to let their own AI call other companies’ MCP servers, so Intercom’s Fin can look something up in Stripe or Shopify, for example. That’s a different feature, often priced separately, and it doesn’t let your AI tools in.

Deacon, built for your coding agent

If you fix bugs with a coding agent, you’ve probably pasted a customer’s complaint into it by hand. Deacon’s MCP server lets the agent read it straight from Deacon, in the customer’s own words, from your terminal or editor.

Here’s what you can ask it once it’s connected.

• “What broke this week?” It reads the week’s bug reports, grouped by topic, with the page each customer was on and the JavaScript errors your site logged. Then it looks for the cause in your code and shows you the edit. Once you say yes, it marks the topic done, and the topic comes back if more people report it.
• “Anything reported since yesterday’s release?” It checks the chats since you shipped, so a bad release doesn’t wait for Monday.
• “Which errors hit the most visits?” Plenty of people hit a bug and leave without a word. Deacon’s script logs those errors too, so your agent can fix the ones nobody reported.
• “Deacon couldn’t answer this. Read the code and suggest an answer.” Your agent writes the answer from how your product actually works. It waits in Conversations, and no customer sees it until you save it.
• “Who reported this?” Once it’s fixed, it brings up the chats behind the bug, with the email each person left, so you can tell them.

It can also tell you which feature people ask for most, read any chat in full and check whether what Deacon knows still matches your code after a big change.

Marking a topic done and suggesting an answer are the only things it changes in Deacon, and it can’t reply to anyone. Your agent is reading what strangers typed, so we left sending out. If you want your own code to reply to customers, Deacon’s API does that. Only an owner on your team can connect an agent, and most sign in through your browser with no key to copy. It’s on every plan, and your agent’s reads don’t count towards your plan’s answers.

Helpdesks with an MCP server

If your team already works in one of these helpdesks, its server comes with it.

Intercom

Intercom’s MCP server reads conversations, contacts, companies and Help Center articles. It can write Help Center articles and internal notes, but it can’t send a reply a customer sees. Workspaces with Fin Operator get five more tools when they sign in rather than use a token.

You sign in through your browser or use an access token. Its connector is in Claude’s directory, and its docs and integration page also name ChatGPT, Claude Code, Cursor, VS Code and Windsurf. Neither its docs nor its pricing page says which plans include it. Where your workspace is hosted matters too. Its developer docs list servers for US and EU workspaces, its integration page says US only, and Australian workspaces aren’t supported yet. A second server lets another AI agent ask Fin a question, and Intercom alternatives for startups has what Intercom costs a startup.

Good for pulling a customer’s history into Claude, not for working the queue.

HubSpot

HubSpot’s remote MCP server has been on every HubSpot account since April 2026. It reads and edits CRM records, tickets included, and HubSpot asks you to confirm before it creates or edits one. It also reads helpdesk conversations from chat, email, WhatsApp and other channels, but it lists no tool for replying in them. If your account has Sensitive Data switched on, it can’t read conversations at all.

You always sign in. HubSpot’s own connectors for Claude, ChatGPT and Copilot run on it, and any other AI tool needs an MCP connector made in your HubSpot account first. HubSpot’s second, local server is for building HubSpot apps, not for reading your CRM. HubSpot vs Intercom sets the two helpdesks side by side.

Good for keeping HubSpot tickets and records up to date from Claude or ChatGPT.

Freshdesk

Freshdesk’s MCP server has been out since 10 September 2026, at your own Freshdesk address. Its 37 tools cover most of a support person’s day, from reading tickets and articles to creating and updating tickets, replying to customers and managing contacts.

There’s no sign-in. Each person connects with their own API key and can only do what their role allows. It’s on Growth, Pro and Enterprise, and every call counts as one action, read or write. The plans include 1,200, 6,000 and 12,000 actions, after which they cost $15 for 1,000. Freshdesk isn’t clear how often that allowance resets. Its pricing table says yearly, the line above it says monthly, and its FAQ says each billing cycle. Freshdesk pricing has what each plan costs.

Good for an AI tool that works tickets end to end, if you watch the action count.

Front

Front’s MCP server reads conversations, messages, contacts, accounts and message templates. It can draft and send messages, assign, tag, archive and comment, though sending needs a permission of its own, which an admin can hold back.

It’s on Starter, Professional and Enterprise, with 100, 200 or 300 tool calls a minute for each teammate. You sign in, but from any AI tool outside Claude’s and ChatGPT’s directories an admin has to make a developer app first. Front pricing goes through the plans.

Good for a team that wants AI to send replies, with an admin deciding who may.

Help Scout

Help Scout’s MCP server is read-only for anyone connecting now. It searches conversations, customers, saved replies, workflows and reports, and your Docs articles if you add a Docs API key. Help Scout says it can’t reply to customers, add notes, change tags or edit Docs, and that write actions are on its roadmap with no date yet.

Each person signs in with their own Help Scout login. It’s on Standard, Plus and Pro, the plans with API access, and the report tools need Plus or Pro. Help Scout pricing has what each plan costs.

Good for searching conversations and pulling reports, not for acting on them.

Gorgias

Gorgias’s MCP server has more than 160 tools, covering tickets, customers, macros, rules, its AI Agent’s settings, the Help Center and reports. Many of them change things, and some send messages to customers, such as replying to a ticket and closing it.

You sign in through your browser, and each tool needs a minimum Gorgias role. It’s on every helpdesk plan at no extra cost, and you don’t need its AI Agent. Gorgias opened it as a beta in May 2026, and on 29 October it became an official connector in Claude and ChatGPT. Gorgias pricing covers the plans.

Good for a shop that wants Claude or ChatGPT to run most of its helpdesk.

Zoho Desk

Zoho Desk has two ways in. The quick one is Zoho’s ready-made Zoho Desk connector in Claude’s directory, which you add by signing in. Its 41 tools include sending a reply.

The other is Zoho MCP, where you build your own server and pick tools from any Zoho app, hundreds of Desk’s among them. For support that means reading tickets, threads and a contact’s history, creating, updating and closing tickets, drafting and sending email replies and writing knowledge base articles. The server’s URL carries a key, so Zoho says to treat it like a password, and you approve each Zoho app you add by signing in. Zoho says it’s free for every Zoho user, within each app’s existing API limits. Zoho Desk pricing has Desk’s own plans.

Good for a team already paying for Zoho, since it costs nothing more.

Plain

Plain’s MCP server reads threads, customers, accounts, tasks, knowledge gaps and the Help Center. It can reply to a thread, assign it, mark it done, add labels and notes and update Help Center articles, and a reply goes out as though you wrote it.

Each person signs in with their own Plain account, since there’s no shared key. Plain’s pricing page lists it on Foundation, Horizon and Frontier, and its docs name ChatGPT, Cursor, Claude on the web and Claude Code.

Good for a team where each person works their own threads from their own AI tool.

Pylon

Pylon’s MCP server has about 100 tools, across issues, accounts, contacts, the knowledge base, feature requests, macros and triggers. It can reply to customers, and Pylon warns that a reply goes straight to the customer’s channel, Slack, email or chat, and can’t be unsent.

Each person signs in, at a US or EU address, and needs MCP Access, which Members and Admins have. It’s on Professional, Advanced and Enterprise, and some tools need an add-on as well.

Good for answering across Slack, email and chat from one AI tool, with care over every send.

Shared inboxes with an MCP server

Missive and Drag can send email for you, while Hiver and Gmelius only save drafts for a person to send.

Missive

Missive’s MCP server searches your conversations, with their messages, comments and attachments, and reads contacts, shared labels, teams, canned responses and your calendars. It can add and remove shared labels and mailboxes on a conversation, write drafts, send them, and create, change or delete calendar events. Sending is a permission of its own. Every permission starts ticked when you connect, so untick sending if you want your AI tool to write drafts and you to send them.

You sign in, with no API key, from Claude’s or ChatGPT’s directory, or at Missive’s server address in any other AI tool. It’s off until an admin turns it on, and if you belong to more than one Missive organisation, every one of them has to turn it on before you can connect. It needs the current Productive or Business plan, and legacy plans need an upgrade. Missive warns that write actions happen straight away, without its own confirmation step, so your AI tool’s approval prompt is the only check before an email goes out.

Good for a team that wants AI to triage and draft in the shared inbox, and to send only if you let it.

Hiver

Hiver has two servers, one for Hiver in Gmail and one for Hiver Omni, its own app. Both read the conversations, messages and tags in the shared inboxes you belong to, and admins also get reports on volume, response times, SLAs, CSAT and sentiment. They can save a shared draft, add internal notes and change tags, and Omni’s can also assign a conversation and change its status. Neither can send, and Hiver says the email goes out only when someone on your team opens the draft and sends it.

Each person signs in with their own Hiver account, and Hiver issues no API keys for it. Its guides add both from Claude’s directory, Hiver in Gmail from ChatGPT’s too, and Hiver Omni through ChatGPT’s Developer mode. Hiver’s pricing pages for Omni and Gmail both include it on Growth, Pro and Elite. It only covers email inboxes, so conversations from your other channels stay out of reach.

Good for a team that wants Claude to triage the shared inbox and draft replies, with a person sending every one.

Gmelius

Gmelius’s MCP server lists your shared inboxes, shared labels and personal inboxes, and reads conversations with their messages, tags and notes. It can assign and unassign conversations, set them open, pending or closed, apply, rename and recolour tags, and write internal notes and draft replies. It can’t send. Gmelius says drafts are never sent automatically, and you review and send them yourself from Gmail.

You sign in to Gmelius in your browser, with no API key to copy, and anything it changes shows up under your name. Its guide adds it as a custom connector in Claude and through Developer Mode in ChatGPT, and also covers Cursor, VS Code and Windsurf. Neither the guide nor the pricing page says which plans include it.

Good for a Gmail team that wants AI to sort and assign the queue and leave drafts, not send.

Drag

Drag’s MCP server has 47 tools. It reads email threads, boards, cards, contacts, the knowledge base, response-time figures and WhatsApp conversations. It can reply to a thread, send a new email, send WhatsApp messages and approved templates, move threads between columns, which is how it assigns them to a teammate, and add labels, comments, cards and knowledge base articles. Drag’s MCP page says a reply can wait as a draft for review, but none of the 47 tools in its docs saves a draft in Drag, and its reply tool sends the email.

You connect with your Drag API key. Paste Drag’s server address into Claude or another AI tool and enter the key once on Drag’s connect page, or run Drag’s open-source server yourself with the key in a config file. Drag is in ChatGPT’s plugin directory, and there ChatGPT asks before it sends email or WhatsApp, though you can change how often it asks. Drag says the key grants full access to your workspace, though the AI tool can only see and do what you can. It’s on the Pro plan.

Good for a Gmail team that wants its AI tool to answer on email and WhatsApp from one inbox.

Live chat tools with an MCP server

Two of the three can reply to customers, and Tidio’s only reads its analytics.

Crisp

Crisp’s MCP server reads conversations, contacts, visitors, operators and the knowledge base. It can send messages to customers, add notes, change a conversation’s state, update profiles and write knowledge base articles.

There’s no sign-in. It uses the same token as Crisp’s REST API, and that token works for the whole workspace, not one person. Crisp’s pricing page lists it on Essentials and Plus, not Free or Mini. A separate server, in beta, lets an AI tool search the knowledge Hugo, Crisp’s AI, answers from. If Crisp’s plans don’t suit you, Crisp alternatives has the other options.

Good for a small team happy to share one token across everyone’s AI tools.

LiveChat

LiveChat now sells its plans as Text, so its MCP server is Text’s. It reads chats, transcripts, tickets, customers and reports, and it can send messages in chats and tickets, tag and transfer chats and end them.

Each person signs in with their Text account, and it follows their permissions. LiveChat’s pricing page includes it on all three Text plans, Essential, Growth and Enterprise, and its API page says older LiveChat plans don’t include it. It’s in Claude’s and ChatGPT’s directories. LiveChat pricing covers the Text plans.

Good for a team already on a Text plan, since every one includes it.

Tidio

Tidio’s MCP server reads its Analytics API and nothing else. It gives an AI tool daily figures and details for conversations, tickets, ratings and Lyro, Tidio’s AI agent, and it can’t send messages or change anything.

It connects with the API credentials a project owner or admin makes, and it’s on Plus and Premium only.

Good for asking how last week went, and nothing more.

AI support chatbots with an MCP server

More than half of the AI chatbots we read have one. They mostly let an AI tool manage the bot and read its chats and leads, and a few can reply.

Chatbase

Chatbase’s MCP server lists 28 tools in its docs. It reads and changes agents, sources, conversations and tickets, exports conversations, chats with your Chatbase agent and sends ticket replies the customer receives. It can also pause the AI in a live chat.

You sign in, with no key to paste, and it’s in Claude’s and ChatGPT’s directories. It needs Basic or higher, and on Free and Hobby every tool but one returns an error.

Good for managing your Chatbase agents and tickets from Claude or ChatGPT.

SiteGPT

SiteGPT’s MCP server reads conversations, leads, escalations, analytics and knowledge sources. It can add sources, change settings and reply to a visitor, taking over their chat.

You can approve it in your browser, where you can make it read-only, or give it an API token. A local version also runs from SiteGPT’s command-line tool. It needs Growth or higher.

Good for stepping into live chats from your AI tool, or only reading them if you’d rather.

DocsBot

DocsBot’s Admin MCP reads and changes bots, sources, conversations, questions, leads, reports and webhooks, and no reply tool is documented. You sign in through your browser, and it’s on every plan. Each bot can also be an MCP server of its own that searches its documentation, from the Standard plan.

Good for managing your bots from Claude, not for answering anyone.

Featurebase

Featurebase’s MCP server reaches its feedback boards, helpdesk conversations and tickets, Help Center, users and reports. It can send replies in conversations and tickets, and write posts, articles and contacts. You sign in, with no key to copy, and it’s on every paid plan.

Good for a team whose feedback board and helpdesk both live in Featurebase.

Kapa.ai

Kapa.ai has two. Its platform server reads projects, conversation threads, top questions and the gaps in your docs, and it can add sources, start crawls, write Q&A pairs and set up integrations and API keys, with no tool for replying. You sign in with a Kapa account, and Kapa’s docs don’t say which plan it needs. A hosted server for each project lets developers search your docs from their own AI tools.

Good for finding the gaps in your docs from your AI tool.

YourGPT

YourGPT’s platform server reads analytics and changes projects, training, contacts and who’s on your team, and it can start a chat with your bot and read past ones. It connects with an API token, and it’s on Professional, Advanced, Agency and Enterprise. A second server only searches a bot’s knowledge base.

Good for running a YourGPT project from your AI tool, not for replying.

CustomGPT.ai

CustomGPT.ai hosts an MCP server for every agent at no extra cost. Out of the box it can only ask your CustomGPT.ai agent a question, and you can allow more, such as reading conversations and analytics, changing settings or syncing knowledge. It connects with a token in the server URL.

Good for letting other AI tools ask your CustomGPT.ai agent a question.

Wonderchat

Wonderchat’s llms.txt describes an MCP server that lets AI tools such as Claude and ChatGPT “work with Wonderchat directly”, and its pricing page lists “API and MCP access” from Basic. Its docs have no page on it, so we can’t say what it reads or changes, or how you connect. The one MCP page in its docs is for a connector that lets its own bot call other servers, from the Scale plan. Wonderchat alternatives compares it with chatbots that do document their servers.

Zendesk’s MCP server isn’t out yet

Zendesk announced an MCP server in May 2026, to “connect Zendesk tickets, knowledge, and other data to external AI systems”, with early access that summer. On 2 November 2026 Zendesk’s developer docs and help centre had no page on using it, and its list of early access programmes didn’t include it. The one programme there that uses an MCP server puts your help centre inside ChatGPT for your customers, which is a different thing. So we can’t yet say what Zendesk’s server does, how you connect or what it costs.

What Zendesk does have is the other direction. Its AI agents and action flows can call other companies’ MCP servers, on every Suite plan and on Support Team and up, and each call uses action credits. The MCP Server app in Zendesk’s Marketplace is made by Swifteq, not Zendesk, and the open-source servers on GitHub work through Zendesk’s API. Zendesk alternatives compares it with tools a small team can start on today.

Built for another job, or none at all

These three have an MCP server, but not one for your support data.

• Voiceflow’s server is for building agents. It edits agents, tools and the knowledge base, publishes, runs test conversations and reads transcripts. You sign in or use a personal key. Voiceflow alternatives looks at it as a chatbot.
• Inkeep offers a server that searches your docs and answers questions about them, and an endpoint that sends a question to one of your agents. Neither reads or replies to support conversations.
• Botpress has a local server that gives a coding agent your project while you build it, and its agents can call other MCP servers. Its changelog mentions a Desk MCP, but we found no page on it. Botpress alternatives compares it with others.

We found no MCP server in the docs of tawk.to, Chatling, Botsonic or Dante AI on 2 November 2026. AI Assist, tawk.to’s own AI, can call other MCP servers, which is the other direction.

Which support MCP server fits you

If your team already works in a helpdesk, use that helpdesk’s own server, since it reaches the same tickets your team does. HubSpot’s, Zoho Desk’s and Gorgias’s come with any plan, most of the others need a particular plan, and Freshdesk counts every call against an allowance. If you want the most an AI tool can do, Zoho Desk through Zoho MCP has hundreds of tools, Gorgias more than 160 and Pylon about 100.

If you want an AI tool to work the queue rather than only read it, pick one that can reply to customers. Of these, 13 can. Hiver’s and Gmelius’s save drafts for a person to send, and Help Scout’s only reads for now. If you’d rather add one from Claude’s or ChatGPT’s own directory than set anything up, Intercom, HubSpot, Front, Gorgias, Zoho Desk, Missive, Hiver, LiveChat and Chatbase are among those listed there.

If you answer support yourself, Deacon is the better fit. It answers your customers from your own docs and shows you what they’re stuck on and what they want next. Its MCP server lets your coding agent read their bug reports beside your code, suggest answers to the questions Deacon couldn’t answer and tell you who to email once it’s fixed.

Questions about support tools’ MCP servers

What is an MCP server in customer support?

It’s how an AI tool such as Claude, ChatGPT or a coding agent connects to your support tool. Once it’s connected, the AI tool can search your conversations, tickets and help articles when you ask it something, and on some servers change them or reply to customers. Most only reach what your own account and role can.

Does Zendesk have an MCP server?

Not one you can set up from Zendesk’s own pages yet. Zendesk announced it in May 2026, with early access that summer, and on 2 November 2026 it had published no docs for it. The MCP Server app in its Marketplace is Swifteq’s. Zendesk’s AI can already call other companies’ MCP servers, which is the other direction.

Which ones can reply to my customers?

On 2 November 2026, Freshdesk’s, Front’s, Gorgias’s, Zoho Desk’s, Plain’s, Pylon’s, Missive’s, Drag’s, Crisp’s, LiveChat’s, Chatbase’s, SiteGPT’s and Featurebase’s could. Front lets an admin hold sending back, and Missive lets you untick sending when you connect. Help Scout’s, Hiver’s, Gmelius’s, Intercom’s, Tidio’s, DocsBot’s, Kapa.ai’s, YourGPT’s, CustomGPT.ai’s and Deacon’s can’t, and HubSpot’s lists no tool for it.

Is it safe to let an AI tool read my support chats?

It can be, with care. Anyone can type anything into a chat, including instructions aimed at an AI tool, so what a customer wrote is something to read, never something to obey. Keep your AI tool’s approval prompts on, so a person says yes before anything changes, and be most careful with a server that can send messages, since a sent reply can’t be taken back. Deacon marks everything a stranger wrote as data and tells your agent not to follow it, and its server can’t reply to anyone. Human in the loop explains the approval step.

Which AI tools can connect to Deacon’s MCP server?

Claude Code, Claude on the web, Codex, Cursor and VS Code sign in through your browser, with no key to copy. ChatGPT adds it the same way on a paid plan. Devin Desktop (Windsurf) connects with a key.

Can I use Deacon alongside Zendesk or Intercom?

Yes, side by side, though they don’t connect by themselves. Deacon has no helpdesk integration. A webhook can tell your own code when a visitor reports a bug or asks for a person, and your code can open the ticket in your helpdesk. Your coding agent can still read the bug reports Deacon hears through its MCP server while your team works in Zendesk or Intercom.