Open-Source Intercom Alternatives in 2026

The reason teams start looking for an Intercom alternative is rarely the product. It's the invoice — and specifically the discovery that the invoice has three axes. You pay per seat, then per AI outcome, then per channel. Add a person and it goes up. Get busier and it goes up. Succeed at automating support and it still goes up, because the automation is metered too. As of September 2026 there is a fourth reason, and it is the one everybody is asking about first: the company is being bought.

Why do teams look for an Intercom alternative?

Because the bill scales on three axes at once. Intercom's seat plans run $29 / $85 / $132 per seat per month billed annually on Essential, Advanced and Expert. On top of that, Fin costs $0.99 per AI outcome, and SMS, WhatsApp, phone and email campaigns are metered separately. Note the wording: Intercom bills per outcome rather than per resolution, charged at most once per conversation, with sales qualification priced at $9.99. Add-ons sit above that again — the Pro bundle from $99/month, Proactive Support Plus at $99/month, and Copilot at $29 per agent per month billed annually, $35 monthly.

Zendesk has the same shape. Published seat prices run $19 per agent per month for Support Team, $55 for Suite Team and $115 for Suite Professional, with Enterprise quoted by sales rather than listed. Then the add-ons: Copilot at $50/agent/month and the Workforce Engagement Bundle at $50/agent/month, with a Contact Center add-on at $83 above those. Stack those on Suite Professional and you are at $215 per agent per month — 87% over the base plan. AI agent usage is billed separately, per automated resolution drawn from a resolution allowance, and since 18 May 2026 resolutions arrive in three tiers: an assisted escalation costs nothing, an unverified contained resolution costs nothing, and only a verified resolution draws down the allowance. Zendesk does not publish the per-resolution rate on its pricing page — and its documentation carries no per-resolution figure — so that number arrives from a salesperson.

None of this makes them bad products. Both are mature, both have reporting a decade deep, and Zendesk has since added Forethought AI agents that run on stacks other than its own. It makes them expensive in a specific way: the more successfully you automate, the more you pay for having done it. It is the same pattern every meter in this industry has — counting something other than the work.

Did Salesforce buy Intercom?

Yes. Salesforce completed its acquisition of Fin — the company formerly named Intercom — on 10 September 2026, at approximately $3.6 billion. The deal was signed on 15 June 2026 and Fin now sits inside Salesforce AI Labs. The naming still trips people up on a call: the company is Fin, the help desk you buy seats on is still branded Intercom, and seat pricing still lives on intercom.com.

Give the product its due, because the numbers behind that price are real. Fin's AI Agent runs on a model suite the company built specifically for customer support, and Salesforce's own release cites deployments resolving an average of 76% of support volume end to end across chat, email, WhatsApp, SMS, phone and Slack. More than 30,000 companies came with the deal. This was not a distressed sale; it was the category leader being bought at a category-leader price.

What changes for a buyer is the planning horizon rather than today's product. Packaging, pricing and roadmap decisions are now made inside a company that has its own agent platform to sell, on a timetable nobody outside it can see, and the thing acquired is specifically the AI agent sitting on top of your support data. If you are signing a multi-year contract this quarter, that is a fair question to put to your account executive and get in writing.

It is also a reasonable moment to ask a more durable question: how much does it cost you to leave, whoever owns the product? That answer does not depend on who signs the acquisition paperwork. It depends on whether your conversation history is something you can take with you — which is a question about export formats, and one worth checking before you need the answer.

What does Intercom actually cost at 1,000 conversations a month?

For a three-agent team on Expert seats, roughly $1,148 a month — $396 in seats plus about $752 in Fin outcomes, if Fin resolves the share Salesforce advertises. On Essential seats the same volume lands near $839. Zendesk Suite Professional with Copilot and the Workforce Engagement Bundle is $645 for three agents before a single automated resolution is counted.

The assumptions matter, so here they are in the open. Three agents, 1,000 conversations a month, annual billing, no channel add-ons. The Fin figure applies Salesforce's own published claim — an average 76% of volume resolved end to end — to those 1,000 conversations, giving 760 outcomes at $0.99. That is a vendor's marketing number being used to size a vendor's bill, which is the fairest way to read it: if Fin performs as advertised, this is what performing as advertised costs.

And note the trade that sits underneath it. A conversation Fin does not resolve costs you nothing in outcomes, but it costs you a human. So the meter runs hardest exactly when the product works best, and the only way to lower that line on the invoice is for the automation to do less. That is not a bug in Intercom's pricing; it is what pricing per outcome means. It is worth deciding whether it is the incentive you want pointed at your support team.

The Zendesk figure is the one with a hole in it. $215 per agent per month is a published number, and $645 for three agents is arithmetic. What is not published is the rate per verified automated resolution — Zendesk's own help centre carries no per-resolution figure — so the automation line on a Zendesk quote cannot be modelled from public information at all. You get it from a salesperson, after you have shown interest.

• 01Intercom Expert — $396 in seats for three agents, plus ~$752 in Fin outcomes at 76% resolution. About $1,148 a month.
• 02Intercom Essential — $87 in seats, same outcome bill. About $839 a month, with a thinner feature set behind it.
• 03Zendesk Suite Professional — $645 for three agents with Copilot and the Workforce Engagement Bundle stacked, before automated resolutions, which are priced privately.
• 04Chatwoot Cloud Business — $117 for three agents, with 500 Captain AI credits included and further credits at $20 per 1,000. Self-hosted Community is $0 in licence.
• 05Munin Cloud Free — €0 per org for all six modules, within 5,000 MCP calls, 250 contacts and 100 MB. Cloud Premium, launching soon, is a flat €99 per org per month rather than per seat.

What is the best open-source Intercom alternative?

Chatwoot is the most mature and the default recommendation — live chat, email, social and WhatsApp in one inbox, and the closest like-for-like replacement. Zammad is better if you need SLA management and ticketing discipline. Rocket.Chat is the pick if internal team chat and customer chat should be one product. Erxes is the only one shipping current mobile SDKs. FreeScout is the lightest. Tiledesk is the bot-first option. Munin is the pick if you want the inbox to share one customer record with your CRM, knowledge base and content, and to be operated by AI agents.

Nine projects in this category present themselves as open-source Intercom alternatives. Two of them — Chaskiq and Papercups — are not maintained, and one, Typebot, is a bot builder rather than a help desk, which is why the shortlist above is shorter than the field. Of the rest, only two have a licence with no carve-out anywhere in it.

The honest framing: most of these replace your inbox. One replaces the reason you had four systems. If it is the CRM side of that you are weighing rather than the inbox, the open-source CRM field compares differently — and mostly on the licence.

Is Chatwoot a good replacement for Intercom?

Yes — Chatwoot is the most mature open-source customer messaging platform available and the closest thing to a drop-in Intercom replacement. It handles live chat, email, social inboxes, and WhatsApp from one unified interface. Expect roughly 2–4 hours to stand up on a VPS; Chatwoot's own deployment page puts the required minimum at 4 GB of RAM, enough for up to 10,000 conversations a day, with 8 GB carrying up to 20,000; how that footprint compares to running a whole platform on one host is its own piece.

The prices are published, which is more than Zendesk manages. Chatwoot Cloud is $0 on Hacker for up to 2 agents and 500 conversations a month, then $19, $39, and $99 per agent per month billed annually on Startups, Business, and Enterprise, with Captain AI credits included per tier — 300, 500, 800 — and further credits at $20 per 1,000. Self-hosted, the Community Edition is free forever, and the supported editions are $19 and $99 per agent per month. Worth reading the shape rather than only the number: self-hosting removes the vendor from your data path, and the paid self-hosted editions are still counted per agent.

One thing to read carefully before you commit: the community edition is MIT, but everything under the enterprise/ directory ships under a separate proprietary, source-available licence. You can read that code, but running those features in production requires a paid Chatwoot subscription and an active agreement. That's open core — a perfectly legitimate business model, and not the same thing as the MIT badge on the front page.

One more thing to check if you are in Europe: Chatwoot's own pricing FAQ states that Chatwoot Cloud runs on AWS with servers in the United States. Self-hosting puts that back under your control entirely, in whichever jurisdiction you choose — which is the answer if data residency is a hard requirement and you are happy to run the box. If you would rather someone else ran it and still needed the data to stay in Europe, that is the case Munin Cloud is built for.

Chatwoot covers WhatsApp, Instagram, and Facebook natively; Munin's four channels are email, chat widget, SMS, and voice. If most of your support arrives through social channels, take Chatwoot. If it arrives through email and chat, and the problem is that your inbox knows nothing about the customer sitting in it, that is the job Munin was built for — one record behind the inbox, MIT throughout, and no enterprise/ directory to buy back.

Chatwoot or Zammad — which should I pick?

Pick Chatwoot if your support is conversational: live chat, social, WhatsApp, people expecting a reply in the widget. Pick Zammad if your support is contractual: tickets, queues, escalation paths, and a number you have promised somebody in writing. The split is not about maturity — both are mature — it is about whether the unit of work is a conversation or a ticket.

The operational bill differs more than the feature lists suggest. Chatwoot is Docker-based and lands in a couple of hours. Zammad wants Ruby on Rails dependencies, Postgres, and 6 GB of RAM before it is comfortable, plus another 4 GB if Elasticsearch runs on the same box — which Zammad calls optional but strongly recommends, because search degrades without it. More moving parts to patch, and a real reason to prefer it only when you need what it does. Licences differ too, and in Zammad's favour: AGPL-3.0 the whole way down with no ee/ directory, against Chatwoot's MIT core plus a proprietary enterprise/ tree.

If you are weighing these two, there is a third question worth asking before you commit to either, because it is the one that is expensive to change later: does the desk need to know who it is talking to? Chatwoot and Zammad both hold a contact record, and both hold it in their own database, separate from whatever your sales side runs. Munin's answer is that the inbox and the CRM are the same table — so a conversation arrives already attached to the customer's history, the campaign they are in, and the knowledge base your team writes. That is the axis Munin competes on, and it is orthogonal to the chat-versus-ticket choice.

What about Zammad and FreeScout?

Zammad is a help desk rather than a chat tool. It takes email, chat, phone, and social, converts them into tickets, and gives you SLA management, reporting, and a knowledge base — the things a support org with contractual obligations actually needs. Its licence is also as clean as copyleft gets: AGPL-3.0 throughout, with no ee/ directory and no cut-down community edition, and self-hosting needs no licence key to get the full feature set. The cost is setup: Ruby on Rails dependencies on a Linux server, a higher bar than the Docker-based options.

FreeScout is a lightweight shared inbox on PHP/Laravel, installable on any LAMP host. It is the only project here that publishes no CPU or RAM minimum at all — its own requirements page says it will run on any system, and the nearest figure anywhere is a wiki note that a $5 server suffices. It's the right answer when what you want is a shared mailbox that isn't Gmail, and nothing more. Its paid modules are unusual in this field for being AGPL-3.0 as well — what you buy is a download key and support, not different terms.

One fact these two share with Chatwoot and Tiledesk, worth stating precisely rather than vaguely: none of them publishes an official MCP server today, and Chatwoot's own extensibility for Captain is custom HTTPS tools rather than MCP. That is not a claim that an agent cannot drive them — a REST API and a community wrapper get you a long way — it is a claim about who is on the hook when the API changes underneath you, and the fuller picture gets a section of its own further down. If it is the cheaper hosted end of the market you are weighing instead, the Freshdesk comparison covers that field and the metering Freshworks has announced for its own MCP interface.

What about Tiledesk?

Tiledesk is an open-source live chat with a visual chatbot builder and human handoff — the closest structural match to what Intercom's Messenger plus Fin actually does. It is written in Node.js, self-hosts with Docker Compose or a Helm chart, and in July 2022 it relicensed to MIT — announced on 20 July, with the deployment repository moving across from GPL-3.0 — collapsing to a single codebase. It is still shipping: its design-studio release feed carries a build dated 9 September 2026. If a bot that deflects the easy questions and passes the rest to a person is the job, this is a serious option.

Credit where it is due, and this is the interesting part: an MIT relicence with one codebase is a genuinely generous move, and it is rarer than the badges on this category's front pages suggest. Most vendors have gone the other way since, in 2025 and 2026 especially.

There is still a line to find, and it is in an unusual place. Tiledesk's own deployment repository notes that Enterprise customers request credentials to pull private, Enterprise-only Docker images. So the carve-out is not a directory you can read and not a per-file header — it is an artefact you cannot see at all. That is the third of the three hiding places worth checking on any project in this category, and Tiledesk is a clean example of it. Not a criticism of the licence, which is real; a note that the licence and the distribution are two different questions.

The scope difference against Munin is the ordinary one. Tiledesk is a conversation surface with bots behind it. Munin is a conversation surface that shares its customer record with a CRM, a knowledge base, a CMS, an outreach tool and analytics, with the whole thing driven through one MCP endpoint — and MIT applies to every part of it, including the images.

Is Rocket.Chat an open-source Intercom alternative?

Partly. Rocket.Chat is a team-chat platform with an Omnichannel module and a Livechat widget, MIT at the core, and it is by a distance the most actively developed project in this comparison — 8.8.1 shipped on 9 September 2026, against 46,000 GitHub stars. The catch is the free tiers: self-hosted Starter stops at 50 users and 100 Monthly Active Contacts, and Community at 100 concurrent users.

Start with the credit, because it is substantial. Rocket.Chat releases weekly, it is one of only two self-hosted projects in this piece with a first-party MCP server — native since 8.8.0, at POST /api/v1/mcp, exposing 11 curated tools by default and 131 with the extended API switched on — and if you want your internal team chat and your customer chat to be the same product rather than two, nothing else open source is close.

The licence is open core, and the mechanism is worth reading precisely rather than from the badge. The root LICENSE is MIT (Expat) for everything except the ee/ and apps/meteor/ee/ directories, which fall under the Rocket.Chat Enterprise Edition licence and require a paid subscription for the correct seat count to run in production. What lives in there matters for this use case specifically: ee/packages/omni-core-ee, ee/packages/omnichannel-services and ee/apps/omnichannel-transcript. Base Omnichannel is in the community code; a good deal of what makes it a support desk is on the far side of that line.

In practice, on the community plan you get one department rather than several, a single business-hours mode, and none of canned responses, the waiting queue, estimated wait time, maximum simultaneous chats, SLA policies and priorities, on-hold handling, or last-agent-preferred routing. Removing the "powered by Rocket.Chat" line from the widget is itself a paid feature. And the number to check before anything else is the MAC cap: Starter includes 100 Monthly Active Contacts, unused ones do not roll over, and Rocket.Chat's own documentation says service may be interrupted if you exceed it without buying a pack. That meter counts the people you talk to, not the people you employ — which is exactly the axis a support desk grows along. Prices are not published either: the pricing page renders Commercial, Government and Defense with "Get pricing" and "Talk to sales", and the self-service Pro plan moved to legacy status on 29 April 2026, closed to new purchases.

So the shape of the decision is clear enough. If one product for internal and external conversation is the goal, take Rocket.Chat and budget for the licensing homework. If what you want is a support desk whose inbox already knows who is writing to it, Munin answers a different question: six modules on one customer record, MIT across the whole repository with no separately licensed directory and no private images, and the MCP tools available on the free tier rather than behind an add-on.

What about Erxes?

Erxes is the most actively maintained of the smaller projects — v3.1.2 shipped on 8 September 2026 — and it is the only project in this comparison with current first-party mobile SDKs: Swift for iOS, Kotlin for Android, plus React Native and Flutter, all updated between June and July 2026. Its licence is AGPLv3 with an added clause forbidding you from hosting it as a SaaS that competes with Erxes.

What it ships is broad: a core of inbox, contacts, products, segments and automation, with Frontline for omnichannel conversations and tickets, Sales, and a Content plugin carrying a knowledge base and help centre. The stack is heavier than Chatwoot's — Node and TypeScript with GraphQL Federation, MongoDB, Redis and Elasticsearch, and React micro-frontends — and the customer-facing surface is the Business Messenger, a script embed of the kind you are replacing.

The licence has one detail worth stating exactly, because it reads differently on paper and in the tree. LICENSE.md declares that plugins ending in -ee under an ee/ directory are Enterprise Edition. But the default branch is now main, and main has no ee/ directory — so the carve-out is declared and currently empty, and everything public is AGPLv3 plus the anti-competition clause. That clause is also why GitHub reports the repository as NOASSERTION rather than AGPL-3.0: it is a real restriction on top of a real copyleft licence, not a badge problem.

One caution on the numbers. Erxes publishes cloud tiers at $0, $49, $99 and $250 a month, but its own pricing page states two contradictory sets of contact limits — the tier cards say 1,000 / 100,000 / unlimited, while the comparison table further down the same page says 5,000 / 25,000 / 200,000. Get the figure in writing before you plan against it.

The shortlisting rule is simple. If your customers reach you from inside your mobile app, Erxes is the one to look at first, and it is not close. If they reach you by email and chat, and the problem you are actually solving is that the inbox knows nothing about the person sitting in it, that is the job Munin was built for — and the difference shows up in the plumbing rather than the feature list, because Erxes federates several databases behind a GraphQL gateway while Munin puts all six modules on one Postgres schema with a single contacts table underneath them.

Which open-source Intercom alternatives should you rule out?

Papercups and Chaskiq, on maintenance, and Typebot on scope. Papercups' marketing site returns a 404 and its hosted app's DNS points at a deleted Heroku dyno; its default branch has had no commit since 10 May 2022. Chaskiq has had a single commit in roughly two years and no release since November 2023. Both still look alive at a glance, because neither repository is archived.

Papercups is the sad one. It is MIT, cleanly, with 6,100 stars and no carve-out of any kind — as permissive as anything here, attached to the deadest project in it. Its README says "maintenance mode", 21 of the 23 repositories in its organisation have not been touched since 2021, and the only thing in the org with recent activity is a community-maintained Flutter client for a backend whose hosted service no longer exists. The code still runs; there is simply nobody behind it.

Chaskiq markets itself directly as an open-source Intercom alternative and, on features, has a fair claim — web messenger, bots, campaigns, help centre, tours, segments. The licence is the problem, and it is not what the badge suggests: AGPL-3.0-or-later plus the Commons Clause, which removes the right to sell, and defines selling to include charging for hosting or support whose value derives substantially from the software. There is a commercial licence that lifts this, mentioned in the README and priced nowhere. Combined with one commit since 2024, it is hard to recommend to anyone.

Typebot is a different case: actively maintained, 10,300 stars, v3.18.0 in August 2026, and honestly licensed under FSL-1.1-Apache-2.0 — a source-available licence that forbids competing uses and then converts each version to Apache-2.0 automatically two years after release. It briefly ran an ee/ directory in 2024 and removed it that September, which is the rarer direction of travel. But it is a chatbot and conversational-form builder, not a help desk: its documentation has no shared team inbox, no agent console and no knowledge base. Pair it with something, or pick something else.

The transferable lesson, since this category is full of the same trap: "not archived" is not "maintained". Check the last commit on the default branch, not the repository's pushed_at timestamp, which counts activity on any branch and made Papercups look two years fresher than it is.

Which open-source Intercom alternatives have mobile SDKs?

Only Erxes, in practice. It ships first-party iOS, Android, React Native and Flutter SDKs, all updated in mid-2026. Chatwoot, Zammad, FreeScout, Tiledesk and Typebot are web-widget only. Rocket.Chat's mobile SDKs are archived — the Android SDK's last commit was in 2018 — and its current iOS and Android apps are for your agents, not for embedding a messenger inside your product.

This is the requirement people forget until migration week, and it is a real one: if you came from Intercom, its mobile SDK may be carrying a meaningful share of your inbound, and no amount of licence comparison helps if the replacement cannot receive it. Worth counting the traffic before you shortlist, not after.

Munin's position here is a scope statement rather than a claim: Munin ships an embeddable chat widget, an email channel, SMS and voice, and does not ship iOS or Android SDKs. If a native in-app messenger is a hard requirement this quarter, Erxes is the shortlist.

What Munin does ship on this axis is the layer underneath, and it is the part that usually causes the trouble later. Every channel writes to the same conversation and the same contact row, so a widget chat, an email and an SMS from the same person are one history rather than three records to reconcile. A mobile app can post into a conversation today through the widget API or the REST endpoint without waiting on an SDK release, and because the whole surface is MCP tools, the agent handling that conversation reaches the CRM, the knowledge base and the outreach campaign in the same pass. The SDK is a delivery mechanism; the customer record is the thing that is expensive to get wrong.

Which of these can an AI agent actually drive?

All of them, to some degree — most publish a REST API, and there is a community MCP wrapper on GitHub for nearly every project here. Three publish a first-party MCP server: Intercom, with 14 tools at mcp.intercom.com and now an EU endpoint at mcp.eu.intercom.com; Rocket.Chat, native since 8.8.0 with 11 curated tools or 131 extended; and Munin, with 240. Chatwoot, Zammad, FreeScout, Tiledesk, Erxes and Chaskiq publish none.

The existence of an endpoint has stopped being the interesting question. Most of this field will have one within a year, and a community wrapper over a REST API genuinely does get an agent a long way. The two questions that survive are what sits behind the endpoint, and what it costs to reach it.

On the first: Intercom's 14 tools cover the help desk that Intercom is, which is the right scope for that product. Rocket.Chat's server reaches its chat surface. Munin's 240 tools reach across all six modules, which is the difference between an agent that closes a ticket and an agent that closes the ticket, updates the contact record, drafts the knowledge-base article that stops the next one, and files the follow-up for a human to approve — in one pass, against one customer row, without crossing a system boundary halfway through.

On the second: Rocket.Chat's MCP server requires the paid AI add-on, which is quoted by sales, plus an access-mcp permission and an admin switch, and it is off by default, marked Alpha, and unavailable on Community or Starter — so the free self-hosted tiers do not have it. Munin's catalogue is the primary surface rather than an add-on to it, identical on Cloud Free and self-hosted, and the REST API is the thing that came second.

One correction worth making, because the marketplaces muddle it: Zendesk publishes an MCP client, not a server. The app listed as "MCP Server" in its marketplace is built by Swifteq, a third-party partner.

What does it actually cost to self-host a support desk?

The licence is the free part, and it is the part people quote. The real bill is three lines: a server, the model tokens your AI actually burns, and somebody's time to patch it. Chatwoot Community, Zammad, FreeScout, Tiledesk and Munin all cost $0 in licence fees self-hosted. None of them costs $0 to run.

The server is the smallest of the three and the easiest to size — a single modest VPS carries any of these at small-team volume, and every vendor above publishes its RAM floor. Check the current rate on your provider's own page rather than a comparison table; the European hosts repriced during 2026 and most third-party figures on the web are stale.

The model tokens are the line that replaces the meter you left behind, and it is worth being exact about what changes. Self-hosting does not make AI support free. It makes the AI cost yours — you pay your model provider directly, at your provider's rate, for the tokens you actually spend. What disappears is the layer between: no $0.99 per outcome, no resolution allowance, no per-1,000-action bundle, and no vendor margin on top of the inference. Whether that is cheaper depends on your volume and your model, and at low volume with heavy automation it usually is by a wide margin.

The third line is the honest one: somebody has to own the box. Backups, upgrades, TLS, the Postgres that fills up. If nobody on your team wants that job, hosted is the right answer and the licence was never the point. That is why Munin ships both from the same repository — Munin Cloud is EU-hosted with a free tier, self-hosting is docker compose up under MIT, and the code is identical either way, so the decision is reversible in both directions.

Is there a free open-source Intercom alternative?

Yes, several, and the differences are in the caps rather than the price. Chatwoot Cloud's Hacker tier is $0 for up to 2 agents and 500 conversations a month. Self-hosted, Chatwoot Community, Zammad and FreeScout's core are free forever under their own licences. Rocket.Chat's Starter tier is free but capped at 100 Monthly Active Contacts. Munin Cloud Free is €0 per org across all six modules.

Read the caps before you read the licence. Chatwoot Hacker gives you live chat only, two agents and 30 days of retention, which is a trial in everything but name; Chatwoot Community self-hosted has no agent cap at all, but Captain AI, voice, SSO and roles sit in the paid enterprise/ tree. Zammad is the cleanest of the free options — AGPL-3.0 the whole way down, no enterprise directory, no licence key, and nothing withheld from the self-hosted build. You pay in RAM rather than money: budget 6 GB, plus 4 more if Elasticsearch shares the box.

FreeScout is the lightest thing here that still runs a real shared inbox, and its core genuinely is AGPL-3.0 and free. The catch is the module store: tags, saved replies, workflows, the API, the knowledge base and reporting are 71 separately-purchased lifetime licences ranging from roughly €3 to €27 each, sold per instance rather than per seat — and the storefront localises its currency by visitor region, so confirm what yours shows before you budget. Unusually for this field, the modules themselves are AGPL-3.0: what you buy is a download key and support, not different terms.

Munin Cloud Free is €0 per org per month rather than per seat, and it is the whole platform rather than a cut-down edition — conversations, CRM, knowledge base, CMS, outreach and analytics, inside 5,000 MCP calls, 250 contacts and 100 MB. Above that, Cloud Premium is a flat €99 per org per month and is launching soon. Self-hosted, Munin is MIT end to end: the free build and the paid build are the same software, with nothing held back in a separately licensed tree.

How does AI pricing compare across these?

Four different meters, and they are not really comparable on price — they are comparable on who carries the risk. Intercom charges $0.99 per Fin outcome. Zendesk draws down a resolution allowance at a rate it does not publish. Chatwoot sells Captain AI credits at $20 per 1,000. Zammad meters AI calls at €0.03 each. Self-host and you pay your model provider directly, at whatever a token costs that week.

The unit matters more than the number. An outcome-priced meter charges you when the automation works, so the bill rises exactly as your deflection rate improves. Intercom is explicit that a resolution, a procedure handoff and a disqualification each count as one outcome at $0.99, with a sales qualification at $9.99, capped at one outcome per conversation. Run Fin on somebody else's help desk — Salesforce, HubSpot, Freshworks, Zoho, Front or Gorgias — and there is no seat cost at all, but a 50-outcome monthly minimum applies.

Zendesk's tiering, introduced on 18 May 2026, is a genuine improvement on a flat count: an assisted escalation is free, a contained resolution that fails verification is free, and only a verified resolution draws down your allowance. What it does not include is a published rate. That comes from a salesperson, which means your number depends partly on how well you negotiate rather than only on what the work costs.

Zammad's €0.03 per AI call prices an operation rather than an outcome, which makes it the fairest metered model in this field: summarising a ticket costs the same whether or not the customer went away happy. FreeScout went further in September 2026 and shipped its AI Integration Module free — the only free module in a catalogue of 71 paid ones.

Munin does not meter AI, because Munin does not sell you the inference. You point Claude, ChatGPT, Cursor or a runner you wrote yourself at the MCP endpoint, and you pay your model provider directly for the tokens you spend. Munin Cloud counts MCP calls against a plan allowance — 5,000 a month on the free tier — and the paid plan, launching soon, is a flat €99 per org per month rather than a per-seat or per-resolution charge. Self-hosted there is no meter at all. Getting better at automation never raises the invoice.

How do I migrate off Intercom?

Contacts and companies move easily; conversation transcripts are the hard part. Intercom's CSV and dataset exports are explicitly reporting metadata only — its own help centre states that conversation and message content is not included. The bulk UI export gives you PDF or text, 200–300 conversations at a time. For real transcripts at volume you have to go through the Conversations API, and the Export API caps each request at a 90-day window.

The practical shape of that, in the order that saves you the afternoon:

1. Contacts and companies first, via CSV or the API. These are the records everything else points at, and loading them first means every foreign key resolves on the way in.
2. Help centre articles next. These export cleanly and are the easiest win — they become your new knowledge base on day one.
3. Conversations last, and decide the depth before you start. Metadata-only via dataset export is fast and gives you volumes, tags and timestamps. Full transcripts mean iterating the Conversations API, 90 days per request, or an S3 JSON export — and note that Intercom's documentation says the S3 export does not contain a transcript either.
4. Decide what history is actually worth moving. Data is available for up to two years. Most teams need the last quarter searchable and the rest archived, and discover that after paying to move all of it.

Into Munin that sequence is crm_import for contacts and companies, kb_import for the help centre, and conv_import for the conversations, in that order — and every one of those has a matching export tool, so the same sequence runs in reverse on the day you want to leave. The HubSpot runbook walks the identical pattern command by command; only the shape of the source export changes.

One detail from Intercom's documentation that is worth sitting with, whichever vendor you pick next: exported conversation data cannot be imported back into a new Intercom workspace. An export you cannot re-import is a backup of a report, not a backup of your data. Ask any vendor — us included — to show you the round trip before you need it.

How is Munin different from a self-hosted help desk?

Munin isn't a help desk with a CRM bolted on. Conversations, CRM, knowledge base, CMS, outreach, and analytics are six modules over one database, and the customer exists in it exactly once. When an email arrives, the sender resolves against the record your outreach campaigns and knowledge base already use. No sync job, no second version of the truth.

That matters most in the moment a support conversation needs to know something. The agent answering it can search the knowledge base, read the customer's history, and check what campaign they're in — not through integrations, but because the six modules were designed against one schema from the start, where the tenant boundary is a database policy rather than something every query has to remember. It also runs backwards: a closed conversation is where the CRM record gets filled in, without anyone retyping what the customer just said. And what your team answered once becomes a knowledge-base article the next person finds instead of another ticket.

The second difference is who operates it. Munin exposes 240 MCP tools at one endpoint, alongside 278 REST endpoints and 62 bundled markdown skills — counted against the live sources on 23 September 2026 — so Claude, ChatGPT, Cursor, or a runner you wrote yourself does the triage, the tagging, the contact extraction, and the knowledge-base drafting directly. Not as a chatbot bolted to the front — as the thing running the desk, with humans approving what matters. Agents propose, humans decide.

And Munin is MIT throughout. No enterprise/ directory, no per-file licence headers, no private images. Leaving is supposed to be easy.

Self-hosted Intercom alternatives compared

Pricing and licence packaging change frequently — confirm on each vendor's own site, including ours. Figures here are indicative and were read off each vendor's own pricing page, repository or documentation on 15 September 2026.

Frequently asked questions

What is the best free open-source alternative to Intercom? Chatwoot for a like-for-like replacement — it's the most mature open-source customer messaging platform and covers chat, email, social, and WhatsApp. Erxes if your customers write to you from inside your mobile app. Tiledesk if a chatbot with human handoff is the main job. Munin if you want the inbox to share one customer record with your CRM, knowledge base, and content, and to be run by AI agents rather than clicked through.

Is Intercom owned by Salesforce? Yes. Salesforce completed the acquisition on 10 September 2026. The deal was signed on 15 June 2026 at approximately $3.6 billion, and Fin — the company formerly named Intercom — now sits inside Salesforce AI Labs. Seat pricing is still published on intercom.com and the help desk is still branded Intercom.

Is the company called Intercom or Fin? Both, and the split trips people up. The company renamed itself Fin in May 2026 and intercom.com/about now redirects to fin.ai. The help desk you buy seats on is still branded Intercom, and seat pricing still lives on intercom.com, while Fin-only pricing lives on fin.ai.

How much does Intercom actually cost? Seats are $29, $85 and $132 per seat per month billed annually on Essential, Advanced and Expert. Fin AI adds $0.99 per outcome, or $9.99 if the outcome is a sales qualification, and SMS, WhatsApp, phone and email campaigns are metered on top. Add-ons start at $99/month, with Copilot at $29 per agent per month annually. Budget well above the headline seat price.

What counts as a $0.99 Fin outcome? Three things, not just resolutions: a resolution, a procedure handoff, and a disqualification. A sales qualification is $9.99. You are charged for at most one outcome per conversation, and escalations, spam and abandoned conversations are not billed. Reading it as "per resolution" understates it.

What does Intercom cost for a small team at 1,000 conversations a month? On a worked example — three agents, annual billing, no channel add-ons — Expert seats come to $396 and Fin outcomes to about $752 if Fin resolves the 76% of volume Salesforce advertises, so roughly $1,148 a month. Essential seats put the same volume near $839. Your figure moves with your resolution rate.

How much does Chatwoot cost? Cloud is $0 on Hacker for 2 agents and 500 conversations a month, then $19, $39, and $99 per agent per month billed annually on Startups, Business, and Enterprise. Self-hosted, the Community Edition is free forever, with Premium Support at $19 and the Enterprise Edition at $99 per agent per month. Captain AI credits are included per tier — 300, 500, 800 — with extra credits at $20 per 1,000.

Is Chatwoot really open source? The community edition is MIT-licensed, so yes for the core. But everything under the enterprise/ directory is proprietary source-available and requires a paid subscription and a valid seat count to run in production. That's an open-core model. Munin has no equivalent directory — the whole repository is MIT, including its container images.

How much RAM does Chatwoot need? 4 GB is the required minimum on Chatwoot's own deployment page, which it says supports up to 10,000 conversations a day; 8 GB supports up to 20,000. Older comparisons citing an 8 GB minimum and 16 GB recommendation are out of date.

Should I choose Chatwoot or Zammad? Chatwoot if the unit of work is a conversation — live chat, social, WhatsApp, replies expected in the widget. Zammad if it is a ticket with an SLA attached, and you need queues, escalation and reporting. Zammad's licence is cleaner (AGPL-3.0 throughout, no ee/ directory, full features self-hosted with no licence key); Chatwoot is markedly easier to stand up and run.

Is Rocket.Chat free for a support desk? Up to a point, and the caps are the thing to check. Self-hosted Starter is free for up to 50 users but includes only 100 Monthly Active Contacts, which do not roll over, and Rocket.Chat's own documentation says service may be interrupted above that without a MAC pack. Community allows 100 concurrent users. Base Omnichannel is in the MIT core, but departments beyond one, canned responses, the waiting queue, SLA policies, on-hold handling and removing the "powered by Rocket.Chat" watermark are all in the proprietary ee/ tree.

Which open-source Intercom alternative has a mobile SDK? Erxes, and effectively only Erxes — first-party Swift, Kotlin, React Native and Flutter SDKs, all updated between June and July 2026. Rocket.Chat's mobile SDKs are archived. Munin approaches it from the other end: it ships a chat widget, email, SMS and voice rather than a native SDK, and every channel writes to one contact row, so a mobile app posts into a conversation through the widget API or the REST endpoint without waiting on an SDK release.

Is Chaskiq open source? Its licence is AGPL-3.0-or-later plus the Commons Clause, which removes the right to sell — including charging for hosting or support whose value derives substantially from the software. A separate commercial licence lifts that, and its terms are not published. The project has also had one commit in roughly two years and no release since November 2023.

Is Papercups still maintained? No. Its marketing site returns a 404, the DNS for its hosted app points at a deleted Heroku dyno, and its default branch has had no commit since 10 May 2022. The repository is not archived and its README says "maintenance mode", which makes it look healthier than it is. The licence is a clean MIT; there is simply nobody behind it.

Is Typebot an Intercom alternative? Not on its own. Typebot is a well-maintained visual chatbot and conversational-form builder under FSL-1.1-Apache-2.0, which converts each version to Apache-2.0 two years after release. It has no shared team inbox, no agent console and no knowledge base, so it replaces the Messenger bot rather than the desk behind it.

Can I export my conversations out of Intercom? Partly, and the detail matters. CSV and dataset exports are reporting metadata only — Intercom's own documentation states message content is not included. Bulk UI exports give PDF or text, 200–300 conversations at a time. Full transcripts at volume require the Conversations API, with the Export API capped at a 90-day window per request, against a two-year retention horizon.

What counts as an automated resolution at Zendesk? Since 18 May 2026, resolutions are tiered. An assisted escalation, where the AI helped but a human finished, is free. A contained resolution that fails verification is also free. Only a verified resolution draws down your allowance, and the waiting period is per channel: email waits 72 hours after the last email, messaging defaults to 2 hours and extends to 72 on request, and voice is assessed at hangup. Zendesk does not publish the rate per resolution, so you get the real one from sales.

Which of these has an official MCP server? Three. Intercom, with 14 tools at mcp.intercom.com and an EU endpoint at mcp.eu.intercom.com. Rocket.Chat, native since 8.8.0, with 11 curated tools or 131 extended — but gated behind its paid AI add-on, off by default, marked Alpha, and unavailable on Community or Starter. And Munin, with 240 tools across all six modules, on the free tier and self-hosted alike. Chatwoot, Zammad, FreeScout, Tiledesk, Erxes and Chaskiq publish none; the servers on GitHub for those are community projects. Zendesk publishes an MCP client — the "MCP Server" app in its marketplace is built by Swifteq, a partner.

Is there an EU-hosted alternative to Intercom? Yes. Munin Cloud is EU-hosted, and self-hosting puts any of these options in whichever jurisdiction you choose. Worth checking before you assume: Chatwoot's pricing FAQ states that Chatwoot Cloud runs on AWS in the United States, so with Chatwoot the EU answer is the self-hosted one. Intercom now offers an EU MCP endpoint, which is not the same thing as EU data residency for the product.

Does self-hosting make AI support free? No — it makes the AI cost yours rather than your help-desk vendor's. You pay your model provider directly for the tokens you spend, with no $0.99 per outcome, no resolution allowance and no margin stacked on the inference. At small-team volume with heavy automation that is usually much cheaper; the trade is that somebody has to run the server.

Can I self-host a help desk with AI support agents? Yes. Munin self-hosts with docker compose up — three containers, Postgres with pgvector plus a backend and a dashboard — and exposes its whole surface as 240 MCP tools, so you point any model at it and the AI cost is whatever your provider charges, not a per-resolution fee to your help-desk vendor.

Which open-source help desk has SLA management? Zammad, and it's the right pick if you're measured on contractual response times. Rocket.Chat has SLA policies too, in its paid ee/ tree. Munin covers a different job: one customer record shared across conversations, CRM, knowledge, and outreach, operated by agents through MCP tools.

Do I need to self-host to escape per-seat pricing? No. Munin Cloud is EU-hosted with a free tier, and the paid plan — a flat €99/month per org rather than per seat — is launching soon. Self-hosting is free forever under MIT if you'd rather run it yourself — same code either way.

If what actually annoys you is that support, sales, and content each hold a different version of the same customer — and that you're metered on automation you could be running yourself — start on Munin Cloud in the EU, or clone the repository and run it on your own machine.

You shouldn't pay more for support the better you get at it.