14 MCP servers for customer feedback in 2026

Short answer. A customer feedback MCP server lets an AI agent (Claude, Claude Code, Cursor, ChatGPT) query your feedback directly instead of you pasting exports. The servers that exist fall into three groups. Support-tool servers (Intercom, Pylon, Plain, and Zendesk's announced one) give the agent raw conversations, one ticket or thread at a time. Feedback-platform servers (Enterpret, Unwrap, Chattermill, BuildBetter, Dovetail, Canny, Featurebase, Productboard) give it analyzed themes, insights, or ideas from that platform's own data. Tracker servers (Linear, Atlassian Rovo for Jira) give it the work. Modem (our product) is a cross-source option: it reads conversations from Slack, Discord, support tools, calls, and issue trackers, groups them into topics with the people, companies, and linked issues behind each one, and exposes that through an MCP server with 16 tools, three of them read-only. If all your feedback lives in one tool, that tool's own server is enough and less to set up.

We build Modem, so weigh its placement with that in mind. Every other entry is limited to what the vendor documents on its own site, checked on 2026-10-07; where we couldn't confirm something, the entry says so.

Raw rows, analyzed themes, or the work

The useful question isn't "which server is best" but "what does the agent get back when it asks?"

Raw rows. A support tool's server returns conversations, contacts, and tickets as they are. The agent does the reading, matching, and counting itself, in every session. Good for "what did this customer tell support?" Weak for "how many customers hit this?"
Analyzed themes. A feedback platform's server returns its own analysis: themes, trends, insights, or board ideas with vote counts. The agent gets a finding instead of a pile of tickets, but only from data that platform already holds.
The work. A tracker's server returns issues, projects, and status. It tells the agent what the team is doing, not what customers said, unless someone already wrote that into the issue.

Most teams end up with one of each. The gap between them (which tickets and Slack threads are the same issue, and is anyone fixing it?) is what a cross-source layer fills.

The short version

Server	Returns	Read or write	Auth	Plan gating (per vendor)	Since
Modem	Topics across connected sources, with messages, people, companies, linked issues and PRs	3 read-only tools, 4 agent-run tools, 9 record tools	OAuth, no API key	Beta; agent runs use credits	Not stated
Enterpret MCP	Analyzed feedback, cited answers	Read-only (Enterpret's wording)	Unconfirmed	Unconfirmed	Announced 2025-10-27
Unwrap	Themes, segments, sampled feedback, trends	Read-only, all 12 tools (Unwrap's wording)	OAuth	Unconfirmed	2026-08-28
Chattermill	Insights, summaries, metrics, raw feedback	Unconfirmed	Unconfirmed	Unconfirmed	Unconfirmed
BuildBetter	Calls, signals, account context, PRDs and briefs	Search and retrieval; personas editable	OAuth, or org API key	Explorer and Pro plans	Unconfirmed
Dovetail	Projects, insights, highlights, research data	Read and create	API token; OAuth where the client supports it	Off with the HIPAA add-on	2025-06-30
Canny	Ideas, insights, comments, users, companies	Read and write, 55 to 60+ tools	OAuth, Canny role permissions	Pro and above with Canny Ideas on	Latest update 2026-08-13
Featurebase	Feedback, roadmaps, support conversations, help center, surveys	Read and write	OAuth as a teammate	All paid plans	2026-05-04
Productboard	Specs, comments, entity status	Read and write	OAuth	Beta; Spark workspaces	2026-06-04
Intercom	Conversations, contacts, companies; Help Center articles	Read; read and write for articles	OAuth or bearer token	Region: check	Unconfirmed
Pylon	Issues, accounts, contacts, KB articles, tasks	Read and update	OAuth only	MCP Access role	Unconfirmed
Plain	Threads, customers, tenants, help center	Read and write, 30 tools	OAuth	Unconfirmed	GA 2026-03-03
Linear	Issues, projects, comments, initiatives	Read and write	OAuth 2.1 or API key	Unconfirmed	2025-05-01
Atlassian Rovo	Jira, Confluence, JSM	Read and write	OAuth 2.1 or API token	Rate limits by plan	GA 2026-02-04

"Unconfirmed" means we didn't find it on the vendor's own site as of 2026-10-07, not that it doesn't exist. Zendesk has announced an official MCP server; check whether it has shipped before you plan around it (details under support tools below).

Cross-source

1. Modem

Modem is for software teams whose customer feedback is spread across Slack, Discord, support tools, sales calls, and issue trackers, and who want their AI agents to answer from all of it at once. It reads conversations from the sources you connect (including Intercom, Zendesk, Pylon, Plain, Canny, Gong, Fathom, Linear, Jira, and GitHub), groups reports of the same problem into topics by meaning, and links each topic to the people and companies who raised it and to the issues and pull requests that address it.

Modem's MCP server is at https://mcp.modem.dev/mcp, signs in with OAuth (no API key to mint or store), and is in beta. It exposes 16 tools (tool reference):

Three read-only tools. search_modem answers a plain-language question with a short answer plus the matching rows, flags partial or truncated results, and spends no agent credits. modem_docs reads Modem's public docs. modem_skills tells the client which tool to use when.
Four agent-run tools. modem_agent_invoke, modem_agent_get_run, modem_agent_send_message, and modem_agent_cancel_run start and manage a full Modem agent run, asynchronously, using agent credits. The run can use your organization's connected integrations, for example filing a Linear issue if Linear is connected, or handing a fix to a coding agent when you ask.
Nine record tools that edit topics, people, and companies as the signed-in user, under that user's Modem role. Merges can't be undone.

A token with only the data:read scope sees just the three read-only tools. Agent turns started over MCP have no approval prompt, so a run acts on what you ask. Calls are limited to 20 per minute per organization per tool.

Register it in Claude Code with claude mcp add --transport http modem https://mcp.modem.dev/mcp; setup for Cursor, VS Code, Codex and others is in the docs. In ChatGPT, install the Modem plugin and pick it with @ in a chat.

Why it works this way. An agent can already reach Slack, Intercom, and Linear one server at a time, but what comes back is raw threads and duplicate reports, and the agent pays for every token it reads while piecing them together. Modem groups the feedback into topics when it arrives, so the agent reads findings with the evidence attached. The read-only search is separate from the paid agent, under its own scope, so the everyday question costs no credits and changes nothing.

Limits. Modem answers only from sources you've connected and from what their backfill windows imported (30 days for most). It doesn't read source code. It has no surveys, NPS, app-store reviews, or in-app widget, and no public voting board.

When Modem is the wrong pick: if a single tool holds almost all your feedback, that tool's own server is enough (Intercom's for an Intercom inbox, Canny's for a Canny board, Linear's if feedback already arrives as Linear issues). If you need survey, NPS, or review analytics at scale, a feedback platform such as Enterpret, Unwrap, or Chattermill fits better. If customers need a public board to vote on, use Canny or Featurebase.

Feedback platforms: analyzed themes

2. Enterpret MCP

Enterpret's MCP server, first called Wisdom MCP and renamed Enterpret MCP, was announced with Enterpret 2.0 on 2025-10-27; a May 2026 upgrade added citations on every answer. Its help center lists four tools (get_organization_details, get_schema, execute_cypher_query, search_knowledge_graph) and names Claude, Claude Code, Cursor, ChatGPT, Notion, and Glean as clients. Enterpret describes it as serving an analyzed view, read-only.

Where it fits: teams already on Enterpret who want its analysis and cited answers inside their AI tools. Enterpret doesn't publish prices; expect a sales conversation.

3. Unwrap

Unwrap announced its MCP server on 2026-08-28. It exposes the same twelve tools as the in-app Unwrap Assistant (list views, search themes and segments, surface the most actionable issues, sample feedback entries, chart over time, run a custom analysis over a sample), and Unwrap says all twelve are read-only. Setup covers Claude, Claude Code, ChatGPT, and Cursor with a one-time OAuth sign-in.

Where it fits: Unwrap customers who want themes and trends in chat. See Unwrap's pricing page for plans.

4. Chattermill

Chattermill's MCP server lets agents such as Claude, Cursor, Codex, and ChatGPT query feedback insights, summaries, metrics, and raw feedback in plain language. We didn't confirm whether it has write tools or how it authenticates.

Where it fits: teams analyzing surveys, reviews, and support at volume in Chattermill. Chattermill prices by feedback volume rather than seats and doesn't publish a list price.

5. BuildBetter

BuildBetter's MCP server is hosted at mcp.buildbetter.app/mcp. Interactive clients sign in with OAuth; unattended automation can use an organization API key in a header. Tools cover searching calls, querying signals, contact and account context, and retrieving PRDs and customer briefs, and synthetic personas can be managed through it. Per BuildBetter's pricing page, MCP access starts at the Explorer plan and isn't on Hobby or Starter.

Where it fits: teams whose richest feedback is recorded calls, analyzed in BuildBetter.

6. Dovetail

Dovetail's MCP server launched on 2025-06-30 at https://dovetail.com/api/mcp, with a self-hosted local option for clients that only support stdio. Tools read projects, insights, highlights, and research data, and can create projects, docs, data, highlights, and channel data. It authenticates with an API token, or OAuth for clients that support it, and Dovetail also lists connectors in the Claude, ChatGPT, Figma, and Microsoft marketplaces. On workspaces with the HIPAA add-on, API keys are off by default, which also disables MCP.

Where it fits: research-led teams, and teams using Dovetail Channels to analyze Zendesk tickets and public Slack channels alongside interviews.

Feedback boards: ideas and requests

7. Canny

Canny's MCP server (https://api.canny.io/api/mcp/v1, OAuth) is the largest in this list, at more than 55 tools per Canny's June changelog and 60+ per its help center. It reads ideas, insights, comments, users, companies, segments, views, and changelog entries, and writes too: create, update, and move ideas, create insights and comments, link HubSpot or Salesforce opportunities, and send items to Autopilot. Tools follow the person's Canny role; merging and deleting are Manager or Owner only. Canny's help center says a Pro or Business plan with Canny Ideas turned on is required, and its changelog says Pro and above; check the article for your plan. See Canny's pricing.

Where it fits: teams running a Canny board who want to triage and update it from their AI tools. Feedback that isn't in Canny isn't visible to it.

8. Featurebase

Featurebase's MCP server, introduced on 2026-05-04, connects Claude, ChatGPT, Cursor, VS Code, Claude Code, and Codex. You sign in as a Featurebase teammate with your own permissions, and agents can read, create, and update feedback, roadmaps, support conversations, customer data, updates, help center content, reports, and surveys. It's on all paid plans.

Where it fits: teams using Featurebase for both a feedback board and support, who want one server for both.

9. Productboard

Productboard's MCP server has been in beta since 2026-06-04 and is included with Spark-enabled workspaces at no extra cost. It authenticates with OAuth at https://mcp.productboard.com. Tools find and read specs, propose and apply edits, read and post comments, and update entity status; since 2026-08-12 coding agents can also create and update Productboard entities, and writes appear in version history. We didn't confirm whether customer notes and insights are exposed, as opposed to specs.

Where it fits: product teams writing specs in Productboard Spark who want coding agents to read and update them.

Support tools: raw conversations

10. Intercom

Intercom's official MCP server is hosted at https://mcp.intercom.com/mcp with 14 tools. It reads conversations, contacts, and companies (search uses a query syntax where you set the object type), and it reads and writes Help Center articles, so it isn't read-only; the OAuth consent also mentions adding internal notes to conversations. Sign in with OAuth or a bearer API token. Intercom's sources disagree on which hosting regions are supported (one says US-hosted workspaces only), so check yours before you set it up.

Where it fits: Intercom teams that want support history, and help articles, reachable from AI tools. Modem also ingests Intercom if you want those conversations grouped with other sources.

11. Pylon

Pylon's MCP server at https://mcp.usepylon.com lets AI tools read and update issues, accounts, contacts, knowledge base articles, tasks, projects, and triggers, within what the signed-in user can see. Authentication is OAuth only; Pylon says API keys aren't supported. An admin enables it, and users need the MCP Access role; some tools also need a plan feature.

Where it fits: B2B support teams running Slack Connect channels through Pylon. Modem ingests Pylon too.

12. Plain

Plain's MCP server at https://mcp.plain.com/mcp went generally available on 2026-03-03 with 30 tools. It reads threads, customers, tenants, and the help center, and writes: reply to threads, assign, change priority, add labels, create notes, mark done. You sign in with your Plain account over OAuth, with your own permissions.

Where it fits: teams that run support in Plain and want agents to act on threads, not just read them. Modem ingests Plain as well.

Zendesk: announced; check whether it has shipped

Zendesk announced an official MCP server at Relate 2026 to connect tickets and knowledge to outside AI systems. As of 2026-10-07 we found no setup documentation for a Zendesk-hosted server, so check whether it has shipped. Separately, Zendesk's MCP client (Zendesk calling other servers from its action flows) is generally available, and a third-party Marketplace app by Swifteq bridges Zendesk tickets to Claude, ChatGPT, and Cursor today. Modem ingests Zendesk tickets and makes them searchable over its own MCP server.

Trackers: the work

13. Linear

Linear's MCP server launched on 2025-05-01 and is hosted at https://mcp.linear.app/mcp. It finds, creates, and updates issues, projects, and comments, and since February 2026 also initiatives, milestones, project updates, and labels. It signs in with OAuth 2.1, or takes a Linear API key as a bearer token; a restricted read-only key gives read-only access.

Where it fits: teams whose customer feedback already becomes Linear issues. It works on issues that exist; something upstream still has to turn feedback into them.

14. Atlassian Rovo MCP (Jira)

Atlassian's Rovo MCP server went generally available on 2026-02-04. It reads, creates, and updates across Jira, Confluence, Jira Service Management, Bitbucket, and Loom, with OAuth 2.1 by default or an API token for headless use. Atlassian says standard Jira and Confluence reads and writes through it don't consume Rovo credits today.

Where it fits: Jira and JSM teams, including support escalations that live in JSM.

Adjacent servers that add context, not feedback

Several servers don't hold feedback but make an agent's answer about a customer better. HubSpot's remote MCP server (generally available since 2026-04-13) and Salesforce's hosted MCP servers (Enterprise Edition and above) return account and deal data, with write tools. Slack's MCP server searches and sends messages once a workspace admin approves the client. Sentry returns errors and traces, and PostHog returns product usage (add ?readonly=true to drop its write tools). Notion's hosted server reads and updates pages.

Build your own feedback MCP server

If your feedback sits in your own database or warehouse, you can write a server with the official MCP SDKs. A useful feedback server usually exposes:

search: a plain-language or keyword search over feedback, returning rows with source links.
get: one item with its full text, author, and company.
group or themes: which items describe the same issue, with counts by company.
trends: volume over time for a theme.
create_issue: file into your tracker, behind a confirmation.

The hard parts are not the server. They're everything behind it: matching the same person across tools, grouping reports by meaning, keeping it fresh, and enforcing permissions. Our guide on how to give AI agents customer context walks through those layers.

Read-only, scopes, and permissions

Three patterns show up across these servers:

Read-only by design. Enterpret and Unwrap describe their servers as read-only. Safe to hand to any agent.
Permissions inherited from the user. Canny, Featurebase, Pylon, Plain, and Modem's record tools act as the signed-in person, so an agent can't do more than its human could.
Scope or key controls. Modem's data:read scope limits a token to read-only tools; Linear accepts restricted read-only API keys; PostHog has a read-only URL flag; Slack requires admin approval per client.

Write tools matter most for headless runs. Many clients ask before calling a tool marked destructive, but an agent run started inside a server (such as a Modem agent run over MCP, which has no approval prompt) acts on what you ask.

How to choose

A coding agent fixing a customer bug: it needs who reported it, their words, and whether it's already tracked. A single support server gets the words for one ticket; Modem gets the grouped topic across sources with linked issues and PRs; Linear or Rovo gets the tracked work.
A product agent asking "what should we build?": an analyzed-themes server (Enterpret, Unwrap, Chattermill, BuildBetter, Dovetail) if you're on one of those platforms; a board server (Canny, Featurebase, Productboard) if requests live on a board; Modem if requests are spread across chat, support, and calls.
A support agent answering a customer: the support tool's own server for the conversation and the account's history (Intercom, Pylon, Plain). Check each server's write tools before counting on it to send the reply: Intercom's documents conversation reads and Help Center writes, not customer replies.
Everything already in one tool: use that tool's server. The cross-source layer earns its setup only when the same issue shows up in several places.

For the client side, see our guides to giving Claude Code and Cursor customer context, and the best ChatGPT apps for customer feedback.

FAQ

What is a customer feedback MCP server?

It's a server that speaks the Model Context Protocol and exposes customer feedback as tools an AI agent can call, such as search, get, or summarize. You add it to Claude, Claude Code, Cursor, ChatGPT, or another MCP client, and the agent queries your feedback mid-task instead of relying on what you pasted.

Should my agent get raw feedback or analyzed themes?

Raw conversations are best when the question is about one customer or one ticket. Analyzed themes are best when the question is "how many" or "what's trending." Most agents need both, plus a way to tell which raw reports belong to which theme.

Which customer feedback MCP servers are read-only?

Enterpret and Unwrap describe theirs as read-only. Intercom's is mostly read but writes Help Center articles. Canny, Featurebase, Productboard, Dovetail, Pylon, Plain, Linear, and Atlassian Rovo all have write tools. Modem's search_modem, modem_docs, and modem_skills are read-only, and a data:read token sees only those three.

Does Zendesk have an official MCP server?

Zendesk announced one at Relate 2026. As of 2026-10-07 we found no setup documentation for a Zendesk-hosted server, so check whether it has shipped. A third-party Marketplace app connects Zendesk tickets to AI tools in the meantime.

Does Modem's MCP server file issues without approval?

The MCP server's own tools don't file issues. An agent run started over MCP (modem_agent_invoke) can, using integrations your organization has connected, and those runs have no approval prompt, so the run acts on what you ask. Modem never files on its own; it acts when a person asks or an automation your team built runs.