MCP for Customer Support: Platforms and Permissions (2026)
Last Updated
Published On
MCP for customer support means your support platform runs an MCP server, so the AI client your team already uses (Claude, ChatGPT, Cursor) can read customer context and act on threads through one standard interface, without an integration layer you build and maintain. Which platforms have a server matters less than whose access each connection runs under, because that decides what the AI can see, what it can change and whose name goes on a reply. Plain is the agentic infrastructure for customer work, a B2B customer support and help desk platform, and its MCP server signs each person in with their own Plain account over OAuth, so a connection carries that person's permissions and no one else's. As of October 28, 2026, Plain's MCP server docs list 43 tools (29 read, 14 write).
What is MCP for customer support?
MCP (Model Context Protocol) is an open standard for connecting AI applications to outside systems. A platform runs an MCP server that describes its tools: a name, a description and an input schema for each. An AI client connects, authenticates and calls those tools when a request needs them, so nobody has to write code against the platform's API first. If the protocol itself is new to you, start with our plain-language MCP explainer.
Direction matters when you compare vendors. A server lets AI clients you choose work with the support platform's data. A client lets the platform's own AI call other systems' servers. Zendesk's August 10, 2026 release is a client: admins connect Zendesk to public or private MCP servers and pick which tools its AI features can use in action flows. That release lets Zendesk's AI reach other systems; on its own, it isn't a way for your AI client to reach Zendesk.
MCP sits on top of the API a platform already has. For Plain that's the GraphQL API, which is why API-first infrastructure matters for AI agents: an MCP server can only expose what the API underneath can do. On our evaluation calls, buyers who ask about MCP usually ask about API access in the same breath: they treat MCP as one door into the platform, not the whole point.
Whose access does an MCP connection inherit?
It depends on how the server authenticates, and vendors answer this in a few different ways. On our recent evaluation calls, a question that kept coming up was about access itself: who a connection signs in as, what it can reach, or who has to approve it. It's the right question to ask first, before tool counts.
Platform | How a connection signs in | Whose permissions it carries | Source (checked October 28, 2026) |
|---|---|---|---|
Plain MCP server | OAuth only, each person once per client | The signed-in person's Plain role; replies and actions appear under that user | |
Intercom MCP server | OAuth (recommended) or a Bearer token | Fin Operator tools are "authorized as you" and need OAuth; not stated for the other tools | |
Intercom Fin Agent API MCP server | A workspace Fin Agent API key | The calling service supplies the end user's identity in the request | |
Front MCP server | OAuth 2.1 with PKCE, via a directory connector or an OAuth app a Front admin creates | The authorizing teammate's role, within the read, write and send permissions the admin allows | |
Freshworks MCP server for Freshdesk | API key only | Not stated in source | |
Dynamics 365 Customer Experience MCP Server for Service | Through Microsoft's Agent 365 Tooling Gateway | Existing Dataverse roles and permissions |
These models behave differently when something goes wrong. With per-person OAuth, a connection can't do more than the person who made it, and every action traces to a named user. With a key, the key's own permissions apply to whoever holds it, so who holds it becomes the real access decision. With a gateway, access follows the platform's existing role model, and the gateway is one more component to configure. Front adds a fourth pattern: an admin sets the outer limit on read, write and send, and each teammate stays inside their own role.
What Plain's model means in practice
Plain's MCP server docs say the server "has the same permissions as your user," and that replies to threads "will appear as though they come from your user." The help article adds the part buyers ask about most:
An admin's connection is the admin's alone. Each person who uses the server completes the OAuth flow once in their own client. There is no shared credential that lets several people connect, so an admin connecting Claude doesn't hand admin access to anyone else.
Any role can connect, including Viewer. Viewer seats are unlimited on every Plain plan, so an engineer or account manager can pull support context into their own AI client without a paid seat, and their connection carries a Viewer's permissions.
Every action is attributed. Replies and actions appear under the user who made them, as they do in the app, which ties MCP activity to a real person for audit and attribution.
A shared connection is possible, with a trade-off. If you want one connection for a whole team, Plain's help article describes building a small custom MCP server over the GraphQL API with a read-only machine user key. All activity then runs as the machine user, so you lose per-person attribution in the workspace timeline.
Permission errors are role errors. The docs' troubleshooting step for a permission error is to check the user's role, not an MCP setting.
When you review any support MCP server, ask the vendor four things: does each person sign in, or does a key stand in for them; whose role bounds the connection; which tools write, and can any of them reach a customer; and where does the action show up afterwards. If the docs don't answer one of these, get the answer before you connect production data.
What can a Plain MCP connection actually do?
It can do most of a support person's queue work, under that person's role. Plain's docs list 43 tools across seven areas as of October 28, 2026. Tools marked write take actions in your workspace, and the docs note that most AI clients ask for confirmation before running them. A longer tool list isn't better on its own, since every definition takes room in the client's context window, so compare vendors on the jobs their lists cover.
Area | Read tools | Write tools |
|---|---|---|
Threads | Fetch Threads, Fetch Thread Details, Fetch My Assigned Threads, Search Threads, Fetch Citations (5) | Reply To Thread, Assign Thread, Unassign Thread, Mark Thread As Done, Mark Thread As Todo, Snooze Thread, Change Thread Priority, Add Labels, Create Note (9) |
Tasks | Fetch Tasks, Fetch My Assigned Tasks, Fetch Task Details, Fetch Task Details By Ref (4) | Create Task, Update Task (2) |
Knowledge | Fetch Knowledge Gaps, Fetch Knowledge Gap Details, Search Knowledge Sources, Fetch Knowledge Sources (4) | Reindex Knowledge Source (1) |
Customers | Fetch Customers, Fetch Customer Details, Fetch Customer Threads, Search Customers (4) | none (0) |
Tenants | Fetch Tenants, Fetch Tenant Details, Search Tenants (3) | Upsert Tenant (1) |
Help Center | Get Help Centers, Get Help Center Article Groups, Get Help Center Articles, Get Help Center Article, Get Help Center Article By Slug (5) | Upsert Help Center Article (1) |
User and workspace | Fetch My User, Fetch My Workspace, Fetch User By Email, Fetch Labels (4) | none (0) |
A few details that matter when you plan what to hand over:
Reply To Thread sends on email, Slack and form threads. If your customers reach you on other channels, check the docs before you plan reply automation.
Thread tools return CSAT (rating, sentiment and comment), and Fetch Threads can filter by it.
Fetch Citations and the Knowledge tools show which sources the Customer Agent (formerly Ari), Plain's customer-facing AI agent, answered from, and where it couldn't answer. You can review its sources from your own client.
Assign Thread can assign to a user, a machine user or yourself.
What it doesn't cover: the documented list has no tool for building workflows. For anything outside the list, the GraphQL API is the next place to look. Sidekick, Plain's agent-facing AI assistant, has its own set of integrations; see the tools Sidekick can use, which are separate from what your own client reaches through MCP.
Which customer support platforms have an MCP server (October 2026)?
As of October 28, 2026, these vendors document an MCP server or client on their own pages. This is not a complete list. The official MCP Registry lets you search for any vendor yourself, and ChatForest's review of helpdesk MCP servers covers community servers too.
Plain, vendor-maintained server. Remote, at
https://mcp.plain.com/mcp, OAuth only. 43 documented tools (29 read, 14 write); Reply To Thread sends customer-visible replies on email, Slack and form threads. Plain's announcement: Plain's MCP server.Intercom, vendor-maintained server. Remote, at
https://mcp.intercom.com/mcp(US) andhttps://mcp.eu.intercom.com/mcp(EU); AU-hosted workspaces are not yet supported. 14 tools, plus five Fin Operator tools for OAuth connections where you can use Operator. Its add_internal_note tool "cannot send customer-visible replies," per the developer guide.Intercom Fin Agent API, vendor-maintained server. At
https://api.intercom.io/fin/mcp, with five tools (list_fin_capabilities, ask_fin, get_fin_response, reply_to_fin_conversation, run_fin_procedure). It exposes Fin as a tool for your agent rather than your workspace's conversations and contacts.Freshdesk, vendor-maintained server. The Freshworks MCP server became generally available on September 10, 2026, at
https://<subdomain>.freshdesk.com/mcp(no custom domains), for Freshdesk Standalone 2021 (latest version) and Freshdesk Unified Omni 2026. It lists 37 tools, including replyTicket, and each successful tool call counts as an action against plan allowances and rate limits.Front, vendor-maintained server. Remote, at
https://mcp.frontapp.com/mcp, over OAuth 2.1 with PKCE. Front's docs say it doesn't support Dynamic Client Registration, so your client must send an OAuth client ID and secret, and some popular assistants aren't compatible yet.Dynamics 365 Customer Service, vendor-maintained server. Generally available as an independent offer, with more than 90 service tools, per Microsoft Learn.
Zendesk, client only. Generally available since August 10, 2026, on Suite Growth and above and Support Team and above. Zendesk's AI calls servers you connect; it doesn't expose Zendesk to your AI client.
How do you connect Plain's MCP server to your AI client?
Every client uses the same URL, https://mcp.plain.com/mcp, and the same sign-in: your own Plain account. These steps are from Plain's docs as of October 28, 2026.
ChatGPT: Settings → Apps → Create app, name it "Plain", enter the URL, set authentication to OAuth, then complete the login. You may need to enable Developer Mode first, or ask an admin to allow custom apps.
Claude.ai: Customize → Connectors → ➕ → Add custom connector, named "Plain" with the URL, then click Connect. On a Claude Team plan, an admin may need to add the connector for you.
Cursor: add the server to ~/.cursor/mcp.json, then click Connect next to Plain in Settings → Tools & MCP.
{
"mcpServers": {
"plain": {
"type": "remote",
"url": "https://mcp.plain.com/mcp"
}
}
}Claude Code: use the mcp-remote helper in ~/.claude.json, because Claude Code doesn't refresh OAuth tokens yet. On the next start, a browser window asks you to sign in and pick a workspace.
"mcpServers": {
"plain": {
"command": "npx",
"args": ["-y", "mcp-remote", "https://mcp.plain.com/mcp"],
"enabled": true
}
}FX or another local client: fx mcp add --transport http plain https://mcp.plain.com/mcp, or point any client that supports remote HTTP servers at the URL and complete OAuth when prompted.
If you build your own client, request the openid offline_access scopes. The docs say a token without them gets a 403 Forbidden even after OAuth succeeds. If a tool doesn't appear, disconnect and reconnect, since some clients cache the tool list.
Start read-only, then move to writes once you trust the results:
"Show me my assigned threads, summarize the top five by urgency, and suggest which to handle first." Calls Fetch My Assigned Threads, then Fetch Thread Details. A long queue can overflow the client's context window, so filter by label or priority first.
"Look up this tenant, show its open threads, and tell me whether anything looks like an escalation risk." Calls Search Tenants, Fetch Tenant Details and Fetch Customer Threads. The answer is only as good as the tenant fields you keep in Plain.
"Show me this week's knowledge gaps. For the top one, check our knowledge sources and draft a Help Center article." Calls Fetch Knowledge Gaps and Search Knowledge Sources, then Upsert Help Center Article if you approve the write.
"Find threads about this error, check our issue tracker for a match, and add an internal note to each thread." Needs your tracker's MCP server connected to the same client. Search Threads is text search with optional filters, so check the date range of what comes back.
How do the Customer Agent, Sidekick and your own agents fit with MCP?
Plain's MCP server is for AI clients your team drives. Plain's own agents, and agents you bring, work inside Plain on the same queue, and each has a different boundary.
The Customer Agent replies to customers under its own guardrails, and it can connect to MCP servers for data it needs while answering. It is charged per resolution.
Sidekick reads code, error logs and call transcripts, and it can use tools through a custom MCP server. It has no channel to the customer. It hands off to a human or to the Customer Agent, which writes and sends the reply. That boundary is what stops internal material from reaching a customer, which MCP alone doesn't solve: Intercom's MCP connectors page warns that data returned to Fin from a tool may be shared with the customer.
Your own agent connects to Plain through External Agents; the bring-your-own-agent announcement explains the model. n8n built its own AI agent on Plain's API, and its case study says that agent handles 60% of n8n's tickets while volume grew from 100 to 2,000 a week. That 60% is n8n's own agent, not the Customer Agent and not an MCP result.
None of these is free. Per Plain's credits docs, as of October 28, 2026:
Action | Credits |
|---|---|
Customer Agent resolves a thread, and it stays in Done for 7 days | 100 |
Connected agent discussion | 15 |
Connected agent thread assignment | 15 |
Sidekick discussion or workflow action | 15–300+ |
Calls from your own AI client to Plain's MCP server aren't on that list: the credits docs say nothing else in Plain uses credits. For how these pieces fit with Sidekick, the Customer Agent or your own agent, and where MCP sits in the wider stack, see The Agentic Support Stack.
Frequently asked questions
Which helpdesks have an MCP server?
As of October 28, 2026, Plain (43 documented tools, OAuth only), Intercom (a 14-tool server plus five Fin Operator tools, and a separate Fin Agent API server), Freshdesk (the Freshworks MCP server, generally available September 10, 2026), Front and Dynamics 365 Customer Service document vendor-maintained MCP servers. Zendesk documents an MCP client, which works in the other direction. Check the official MCP Registry for any vendor not listed here.
If an admin connects Plain's MCP server, does everyone get admin permissions?
No. Plain's MCP server uses OAuth only, and each person completes the sign-in in their own client, so a connection carries only the permissions of the person who made it. There is no shared credential. A team that wants one shared connection can build a custom server with a read-only machine user key, at the cost of per-person attribution.
Does Plain's MCP server support API key authentication?
No. As of October 28, 2026, Plain's help center says the server at https://mcp.plain.com/mcp supports OAuth only. The documented alternative is a custom MCP server that calls Plain's GraphQL API with a machine user API key.
Can an MCP server send replies to customers?
It depends on the vendor's tools. Plain's Reply To Thread sends on email, Slack and form threads, and the reply appears under the connected user. Freshdesk lists replyTicket. Intercom's MCP server documents add_internal_note, which cannot send customer-visible replies.
Does using Plain's MCP server cost credits?
No. As of October 28, 2026, Plain's credits docs list the Customer Agent, connected agents, Sidekick and AI-selected labels or priority as the only things that use credits. A Plain plan is still required: Foundation is $45 per user per month, billed annually, with a 7-day free trial and no credit card required.
Which AI clients work with Plain's MCP server?
Plain's docs give setup steps for ChatGPT, Claude.ai, Claude Code, Cursor and FX, and any local client that supports remote HTTP servers can connect to https://mcp.plain.com/mcp and complete OAuth.
Plain's MCP server gives the AI client your team already uses 43 documented tools, under each person's own Plain role. Start a 7-day free trial to connect it, or book a demo to see how it fits your support setup.
MCP for customer support means your support platform runs an MCP server, so the AI client your team already uses (Claude, ChatGPT, Cursor) can read customer context and act on threads through one standard interface, without an integration layer you build and maintain. Which platforms have a server matters less than whose access each connection runs under, because that decides what the AI can see, what it can change and whose name goes on a reply. Plain is the agentic infrastructure for customer work, a B2B customer support and help desk platform, and its MCP server signs each person in with their own Plain account over OAuth, so a connection carries that person's permissions and no one else's. As of October 28, 2026, Plain's MCP server docs list 43 tools (29 read, 14 write).
What is MCP for customer support?
MCP (Model Context Protocol) is an open standard for connecting AI applications to outside systems. A platform runs an MCP server that describes its tools: a name, a description and an input schema for each. An AI client connects, authenticates and calls those tools when a request needs them, so nobody has to write code against the platform's API first. If the protocol itself is new to you, start with our plain-language MCP explainer.
Direction matters when you compare vendors. A server lets AI clients you choose work with the support platform's data. A client lets the platform's own AI call other systems' servers. Zendesk's August 10, 2026 release is a client: admins connect Zendesk to public or private MCP servers and pick which tools its AI features can use in action flows. That release lets Zendesk's AI reach other systems; on its own, it isn't a way for your AI client to reach Zendesk.
MCP sits on top of the API a platform already has. For Plain that's the GraphQL API, which is why API-first infrastructure matters for AI agents: an MCP server can only expose what the API underneath can do. On our evaluation calls, buyers who ask about MCP usually ask about API access in the same breath: they treat MCP as one door into the platform, not the whole point.
Whose access does an MCP connection inherit?
It depends on how the server authenticates, and vendors answer this in a few different ways. On our recent evaluation calls, a question that kept coming up was about access itself: who a connection signs in as, what it can reach, or who has to approve it. It's the right question to ask first, before tool counts.
Platform | How a connection signs in | Whose permissions it carries | Source (checked October 28, 2026) |
|---|---|---|---|
Plain MCP server | OAuth only, each person once per client | The signed-in person's Plain role; replies and actions appear under that user | |
Intercom MCP server | OAuth (recommended) or a Bearer token | Fin Operator tools are "authorized as you" and need OAuth; not stated for the other tools | |
Intercom Fin Agent API MCP server | A workspace Fin Agent API key | The calling service supplies the end user's identity in the request | |
Front MCP server | OAuth 2.1 with PKCE, via a directory connector or an OAuth app a Front admin creates | The authorizing teammate's role, within the read, write and send permissions the admin allows | |
Freshworks MCP server for Freshdesk | API key only | Not stated in source | |
Dynamics 365 Customer Experience MCP Server for Service | Through Microsoft's Agent 365 Tooling Gateway | Existing Dataverse roles and permissions |
These models behave differently when something goes wrong. With per-person OAuth, a connection can't do more than the person who made it, and every action traces to a named user. With a key, the key's own permissions apply to whoever holds it, so who holds it becomes the real access decision. With a gateway, access follows the platform's existing role model, and the gateway is one more component to configure. Front adds a fourth pattern: an admin sets the outer limit on read, write and send, and each teammate stays inside their own role.
What Plain's model means in practice
Plain's MCP server docs say the server "has the same permissions as your user," and that replies to threads "will appear as though they come from your user." The help article adds the part buyers ask about most:
An admin's connection is the admin's alone. Each person who uses the server completes the OAuth flow once in their own client. There is no shared credential that lets several people connect, so an admin connecting Claude doesn't hand admin access to anyone else.
Any role can connect, including Viewer. Viewer seats are unlimited on every Plain plan, so an engineer or account manager can pull support context into their own AI client without a paid seat, and their connection carries a Viewer's permissions.
Every action is attributed. Replies and actions appear under the user who made them, as they do in the app, which ties MCP activity to a real person for audit and attribution.
A shared connection is possible, with a trade-off. If you want one connection for a whole team, Plain's help article describes building a small custom MCP server over the GraphQL API with a read-only machine user key. All activity then runs as the machine user, so you lose per-person attribution in the workspace timeline.
Permission errors are role errors. The docs' troubleshooting step for a permission error is to check the user's role, not an MCP setting.
When you review any support MCP server, ask the vendor four things: does each person sign in, or does a key stand in for them; whose role bounds the connection; which tools write, and can any of them reach a customer; and where does the action show up afterwards. If the docs don't answer one of these, get the answer before you connect production data.
What can a Plain MCP connection actually do?
It can do most of a support person's queue work, under that person's role. Plain's docs list 43 tools across seven areas as of October 28, 2026. Tools marked write take actions in your workspace, and the docs note that most AI clients ask for confirmation before running them. A longer tool list isn't better on its own, since every definition takes room in the client's context window, so compare vendors on the jobs their lists cover.
Area | Read tools | Write tools |
|---|---|---|
Threads | Fetch Threads, Fetch Thread Details, Fetch My Assigned Threads, Search Threads, Fetch Citations (5) | Reply To Thread, Assign Thread, Unassign Thread, Mark Thread As Done, Mark Thread As Todo, Snooze Thread, Change Thread Priority, Add Labels, Create Note (9) |
Tasks | Fetch Tasks, Fetch My Assigned Tasks, Fetch Task Details, Fetch Task Details By Ref (4) | Create Task, Update Task (2) |
Knowledge | Fetch Knowledge Gaps, Fetch Knowledge Gap Details, Search Knowledge Sources, Fetch Knowledge Sources (4) | Reindex Knowledge Source (1) |
Customers | Fetch Customers, Fetch Customer Details, Fetch Customer Threads, Search Customers (4) | none (0) |
Tenants | Fetch Tenants, Fetch Tenant Details, Search Tenants (3) | Upsert Tenant (1) |
Help Center | Get Help Centers, Get Help Center Article Groups, Get Help Center Articles, Get Help Center Article, Get Help Center Article By Slug (5) | Upsert Help Center Article (1) |
User and workspace | Fetch My User, Fetch My Workspace, Fetch User By Email, Fetch Labels (4) | none (0) |
A few details that matter when you plan what to hand over:
Reply To Thread sends on email, Slack and form threads. If your customers reach you on other channels, check the docs before you plan reply automation.
Thread tools return CSAT (rating, sentiment and comment), and Fetch Threads can filter by it.
Fetch Citations and the Knowledge tools show which sources the Customer Agent (formerly Ari), Plain's customer-facing AI agent, answered from, and where it couldn't answer. You can review its sources from your own client.
Assign Thread can assign to a user, a machine user or yourself.
What it doesn't cover: the documented list has no tool for building workflows. For anything outside the list, the GraphQL API is the next place to look. Sidekick, Plain's agent-facing AI assistant, has its own set of integrations; see the tools Sidekick can use, which are separate from what your own client reaches through MCP.
Which customer support platforms have an MCP server (October 2026)?
As of October 28, 2026, these vendors document an MCP server or client on their own pages. This is not a complete list. The official MCP Registry lets you search for any vendor yourself, and ChatForest's review of helpdesk MCP servers covers community servers too.
Plain, vendor-maintained server. Remote, at
https://mcp.plain.com/mcp, OAuth only. 43 documented tools (29 read, 14 write); Reply To Thread sends customer-visible replies on email, Slack and form threads. Plain's announcement: Plain's MCP server.Intercom, vendor-maintained server. Remote, at
https://mcp.intercom.com/mcp(US) andhttps://mcp.eu.intercom.com/mcp(EU); AU-hosted workspaces are not yet supported. 14 tools, plus five Fin Operator tools for OAuth connections where you can use Operator. Its add_internal_note tool "cannot send customer-visible replies," per the developer guide.Intercom Fin Agent API, vendor-maintained server. At
https://api.intercom.io/fin/mcp, with five tools (list_fin_capabilities, ask_fin, get_fin_response, reply_to_fin_conversation, run_fin_procedure). It exposes Fin as a tool for your agent rather than your workspace's conversations and contacts.Freshdesk, vendor-maintained server. The Freshworks MCP server became generally available on September 10, 2026, at
https://<subdomain>.freshdesk.com/mcp(no custom domains), for Freshdesk Standalone 2021 (latest version) and Freshdesk Unified Omni 2026. It lists 37 tools, including replyTicket, and each successful tool call counts as an action against plan allowances and rate limits.Front, vendor-maintained server. Remote, at
https://mcp.frontapp.com/mcp, over OAuth 2.1 with PKCE. Front's docs say it doesn't support Dynamic Client Registration, so your client must send an OAuth client ID and secret, and some popular assistants aren't compatible yet.Dynamics 365 Customer Service, vendor-maintained server. Generally available as an independent offer, with more than 90 service tools, per Microsoft Learn.
Zendesk, client only. Generally available since August 10, 2026, on Suite Growth and above and Support Team and above. Zendesk's AI calls servers you connect; it doesn't expose Zendesk to your AI client.
How do you connect Plain's MCP server to your AI client?
Every client uses the same URL, https://mcp.plain.com/mcp, and the same sign-in: your own Plain account. These steps are from Plain's docs as of October 28, 2026.
ChatGPT: Settings → Apps → Create app, name it "Plain", enter the URL, set authentication to OAuth, then complete the login. You may need to enable Developer Mode first, or ask an admin to allow custom apps.
Claude.ai: Customize → Connectors → ➕ → Add custom connector, named "Plain" with the URL, then click Connect. On a Claude Team plan, an admin may need to add the connector for you.
Cursor: add the server to ~/.cursor/mcp.json, then click Connect next to Plain in Settings → Tools & MCP.
{
"mcpServers": {
"plain": {
"type": "remote",
"url": "https://mcp.plain.com/mcp"
}
}
}Claude Code: use the mcp-remote helper in ~/.claude.json, because Claude Code doesn't refresh OAuth tokens yet. On the next start, a browser window asks you to sign in and pick a workspace.
"mcpServers": {
"plain": {
"command": "npx",
"args": ["-y", "mcp-remote", "https://mcp.plain.com/mcp"],
"enabled": true
}
}FX or another local client: fx mcp add --transport http plain https://mcp.plain.com/mcp, or point any client that supports remote HTTP servers at the URL and complete OAuth when prompted.
If you build your own client, request the openid offline_access scopes. The docs say a token without them gets a 403 Forbidden even after OAuth succeeds. If a tool doesn't appear, disconnect and reconnect, since some clients cache the tool list.
Start read-only, then move to writes once you trust the results:
"Show me my assigned threads, summarize the top five by urgency, and suggest which to handle first." Calls Fetch My Assigned Threads, then Fetch Thread Details. A long queue can overflow the client's context window, so filter by label or priority first.
"Look up this tenant, show its open threads, and tell me whether anything looks like an escalation risk." Calls Search Tenants, Fetch Tenant Details and Fetch Customer Threads. The answer is only as good as the tenant fields you keep in Plain.
"Show me this week's knowledge gaps. For the top one, check our knowledge sources and draft a Help Center article." Calls Fetch Knowledge Gaps and Search Knowledge Sources, then Upsert Help Center Article if you approve the write.
"Find threads about this error, check our issue tracker for a match, and add an internal note to each thread." Needs your tracker's MCP server connected to the same client. Search Threads is text search with optional filters, so check the date range of what comes back.
How do the Customer Agent, Sidekick and your own agents fit with MCP?
Plain's MCP server is for AI clients your team drives. Plain's own agents, and agents you bring, work inside Plain on the same queue, and each has a different boundary.
The Customer Agent replies to customers under its own guardrails, and it can connect to MCP servers for data it needs while answering. It is charged per resolution.
Sidekick reads code, error logs and call transcripts, and it can use tools through a custom MCP server. It has no channel to the customer. It hands off to a human or to the Customer Agent, which writes and sends the reply. That boundary is what stops internal material from reaching a customer, which MCP alone doesn't solve: Intercom's MCP connectors page warns that data returned to Fin from a tool may be shared with the customer.
Your own agent connects to Plain through External Agents; the bring-your-own-agent announcement explains the model. n8n built its own AI agent on Plain's API, and its case study says that agent handles 60% of n8n's tickets while volume grew from 100 to 2,000 a week. That 60% is n8n's own agent, not the Customer Agent and not an MCP result.
None of these is free. Per Plain's credits docs, as of October 28, 2026:
Action | Credits |
|---|---|
Customer Agent resolves a thread, and it stays in Done for 7 days | 100 |
Connected agent discussion | 15 |
Connected agent thread assignment | 15 |
Sidekick discussion or workflow action | 15–300+ |
Calls from your own AI client to Plain's MCP server aren't on that list: the credits docs say nothing else in Plain uses credits. For how these pieces fit with Sidekick, the Customer Agent or your own agent, and where MCP sits in the wider stack, see The Agentic Support Stack.
Frequently asked questions
Which helpdesks have an MCP server?
As of October 28, 2026, Plain (43 documented tools, OAuth only), Intercom (a 14-tool server plus five Fin Operator tools, and a separate Fin Agent API server), Freshdesk (the Freshworks MCP server, generally available September 10, 2026), Front and Dynamics 365 Customer Service document vendor-maintained MCP servers. Zendesk documents an MCP client, which works in the other direction. Check the official MCP Registry for any vendor not listed here.
If an admin connects Plain's MCP server, does everyone get admin permissions?
No. Plain's MCP server uses OAuth only, and each person completes the sign-in in their own client, so a connection carries only the permissions of the person who made it. There is no shared credential. A team that wants one shared connection can build a custom server with a read-only machine user key, at the cost of per-person attribution.
Does Plain's MCP server support API key authentication?
No. As of October 28, 2026, Plain's help center says the server at https://mcp.plain.com/mcp supports OAuth only. The documented alternative is a custom MCP server that calls Plain's GraphQL API with a machine user API key.
Can an MCP server send replies to customers?
It depends on the vendor's tools. Plain's Reply To Thread sends on email, Slack and form threads, and the reply appears under the connected user. Freshdesk lists replyTicket. Intercom's MCP server documents add_internal_note, which cannot send customer-visible replies.
Does using Plain's MCP server cost credits?
No. As of October 28, 2026, Plain's credits docs list the Customer Agent, connected agents, Sidekick and AI-selected labels or priority as the only things that use credits. A Plain plan is still required: Foundation is $45 per user per month, billed annually, with a 7-day free trial and no credit card required.
Which AI clients work with Plain's MCP server?
Plain's docs give setup steps for ChatGPT, Claude.ai, Claude Code, Cursor and FX, and any local client that supports remote HTTP servers can connect to https://mcp.plain.com/mcp and complete OAuth.
Plain's MCP server gives the AI client your team already uses 43 documented tools, under each person's own Plain role. Start a 7-day free trial to connect it, or book a demo to see how it fits your support setup.