Open-source alternative to Intercom Fin AI agent

Can an open-source AI agent replace Intercom Fin for our support queue?

Chatwoot's Captain is the open-source option for this job. It is available on the self-hosted Premium plan, and it replaces Fin's function of answering customer queries automatically, not Fin's billing model or its managed AI service. Captain is not included in the Community Edition, which is free but lacks it, so a team that wants an AI agent on its own servers should budget for the Premium plan from the start. The Chatwoot self-hosted plans page lists Captain among the features of that tier, and the plan prices are there as well.

Captain does not arrive with a model or a body of knowledge. Your organization supplies both. You need to choose the language model Captain calls and decide where that model runs, since a hosted model API sends conversation content to a third party while a model you host keeps it inside your environment. You also need a knowledge source, such as your help center articles, product documentation and approved answers, that is current and written for customers. Before cutover, confirm with Chatwoot which model providers Captain supports in your version, what data leaves your network in each configuration, and how answers are reviewed when Captain is unsure.

The cost shape changes in ways you should model before deciding. Fin's charge follows resolved outcomes, so it rises with volume. With Captain, the license cost is per agent per month, and the recurring costs that scale with volume are model inference, the servers that run Chatwoot and its Sidekiq workers, and the staff who maintain the knowledge base. Ask for a quote that separates the Premium license, the model and its usage charges, hosting and the ongoing content work. Run a pilot on a single inbox with a sample of real questions, and measure how many Captain answers a human still has to correct before you commit.

Which Fin capabilities map to Chatwoot Captain and what is missing

This table maps the main Intercom Fin capabilities to Chatwoot Captain, the Chatwoot inbox and Help Center portal, and Zammad where it applies. Read the middle columns for what the documentation describes and the last column for what to verify before you rely on it. Captain is described as automating responses to common queries, but resolution tracking, testing tools and model choice need written confirmation from Chatwoot.

Fin capability	Chatwoot (Captain, inbox, Help Center)	Zammad	What to confirm
Answer generation for common queries	Captain automates responses to common queries, per the Chatwoot README	No AI answer agent found in the researched sources; treat as a gap	Answer quality on your own content and your own policies
Knowledge source use	Confirm which content Captain reads, including whether Help Center articles are indexed	Knowledge base with public and internal answers	Sources Captain uses, and how edits to articles reach the AI
Handoff to human agents	Auto-assignment by agent availability, teams and private notes; Captain handoff rules not documented in the sources	Ticket queues and agent assignment; no AI handoff documented	Handoff triggers, and whether conversation history passes to the agent
Channel coverage	Website live chat, email, Facebook, Instagram, Twitter, WhatsApp, Telegram, Line and SMS in one inbox	Strongest on email and ticket channels; confirm chat and social options	Which channels Captain answers on, not only which channels the inbox receives
Help Center	Help Center portal for articles, FAQs and guides	Knowledge base for public and internal answers	Article migration from Intercom and in-widget search behavior
Reporting and resolution tracking	Conversation, agent, inbox, team and CSAT reports listed in the README	Standard ticket reporting; AI resolution metrics not documented	How Captain resolutions are counted, and whether they can be exported
Multilingual support	Google Translate integration listed in the README	Knowledge base language coverage to confirm	Answer accuracy in each language you serve, with Captain enabled
Testing tools and model choice	Not covered in the researched sources	Not covered in the researched sources	Test tooling before go-live, and which model providers Captain supports

Chatwoot Captain, Premium plan and build-your-own agent compared

Chatwoot offers its Captain AI agent only on paid self-hosted tiers, so the choice is partly a budget decision and partly an operating decision. The three options below differ in licence cost, what you must build yourself, and how identity and service levels are handled.

Captain on Premium Support

Premium Support adds Captain AI, voice calls, custom branding, agent capacity, roles and permissions and priority support to the self-hosted Chatwoot product. The fee is charged per agent per month, so the bill grows with headcount rather than with the number of AI-resolved conversations. Ask Chatwoot which Captain features, model providers and knowledge source types are included, and whether model API usage is billed to you separately from the licence.

Community Edition with a self-built agent

The Community Edition is free and does not include Captain, so teams on it who want an AI first response must build the agent themselves. A typical pattern is a service that reads new conversations through the Chatwoot API, queries your knowledge base, calls a model you host or license, and posts a reply or hands off to an agent. You own the prompts, guardrails, retrieval quality and failure handling, and you need engineering time to keep them current. Expect the work to recur as models and content change.

Enterprise Edition for SSO and SLAs

Choose the Enterprise Edition when the requirement is identity or contractual, not AI. It adds SSO with SAML and SLA policies on top of the per-agent fee. If your agents must sign in through a corporate identity provider and your service levels need to be written into a contract, this tier is the only listed route to both. It does not, on the listed plan description, replace the need to design and test the AI agent, so confirm whether Captain is included in your chosen tier before you buy.

Self-built agent trade-offs

A self-built agent removes the per-agent AI licence but moves cost into hosting, model usage and staff time. Model inference, vector search and a reranking or evaluation service each add infrastructure that must be patched, monitored and backed up next to Chatwoot's own servers. Before committing, define how the agent escalates to a person, how it records what it answered, and who approves changes to the knowledge base. Illustratively, a team that answers a narrow set of order-status questions may find this manageable, while a broad product support queue will need more testing.

Questions to put to Chatwoot

Before you choose, request written answers on four points: whether Captain is bundled in the tier you buy, which model providers it can use and whether you must supply your own keys, how knowledge sources are indexed and refreshed, and how conversation content sent to AI providers is handled under your data terms. Ask for the exact per-agent billing rules, including how agents are counted. These answers determine the total cost, so get them in the quote before you compare totals with Intercom.

What a self-hosted AI agent needs for model hosting and knowledge

Captain generates answers through a language model, and that model is the first decision that determines where customer messages travel. A self-hosted Chatwoot installation keeps the conversation database on your servers, but a model served by an external provider receives the text of each prompt it is sent. A model you host yourself keeps that text inside your network, at the cost of GPU capacity, model operations and patching. Confirm with Chatwoot which model providers Captain supports and how each one is configured, including where keys are entered and stored.

External provider: Prompts and retrieved snippets leave your environment. Review the provider's data retention, training and regional terms, and who in your organization holds the API key.
Self-hosted model: Inference runs in your data center or cloud account, so prompts stay within your control. Budget for accelerator hardware, model serving and capacity planning.
Key ownership: Store provider keys in your secrets tooling, rotate them on a schedule, and restrict which administrators can view or change them.
Recurring cost: Provider usage is typically billed by tokens or requests, so cost grows with conversation volume. Ask any provider for its published pricing and whether a business agreement changes it.
Fallback behavior: Decide what happens when the model is unavailable, so customers are routed to a human agent.

Before approving a design, confirm the supported providers, the configuration keys and the data flows with Chatwoot documentation.

How cost behaves as conversation volume grows on Chatwoot

Fin's spend rises with every resolution, while Chatwoot's Premium spend rises with agent seats and the infrastructure under them. That difference decides how the two bills behave as support volume grows. Intercom charges for Fin per resolved outcome, so a busy month of automated resolutions raises the invoice directly. On a self-hosted Chatwoot stack, the subscription side is priced per agent: the Premium Support plan lists Captain AI at $19 per agent per month on the self-hosted plans page, so adding agents adds a predictable line. Volume itself mainly lands on your servers and on your model provider.

The infrastructure side follows conversation load. Chatwoot's deployment requirements recommend at least 4 CPU cores and 4 GB of RAM for up to 10,000 conversations a day, and 8 cores and 8 GB for up to 20,000, plus at least 1 GB of swap. The same page notes that Sidekiq, the background job processor, can use more than 1 GB of memory on a very active server, so size for peaks rather than averages. Check whether your cloud or data center bill for these hosts is already covered by existing capacity, and ask your platform team how a move to the next tier would be provisioned and who would run it.

Model API usage is the cost to plan most carefully, because it recurs and you pay it directly to the model provider, not to Chatwoot. Captain calls a language model for every automated answer, so its cost depends on how many conversations it touches, how much knowledge it retrieves for each answer, and which model you choose. No figures for model usage are supplied for this page, so ask your team or DeployOpen for a written estimate built from your own monthly conversation counts, share of automated answers and expected tokens per exchange. Ask for the same estimate at two or three volume levels so you can see the curve before you commit.

Moving from Fin to a self-hosted Captain rollout

This sequence moves AI-assisted replies from Intercom Fin to Chatwoot Captain in a controlled order, so each stage has a check before the next one starts. The result is a live, self-hosted Captain on a pilot inbox, validated against Fin, and a cutover plan the organization has signed off in writing.

Confirm the plan, edition and model access
Before any installation, confirm which Chatwoot edition includes Captain. The Chatwoot self-hosted plans page lists Captain AI under Premium Support, so the Premium plan is the starting point, and the Community Edition is not sufficient for this section. Confirm with Chatwoot which model providers Captain can call, and who holds the API keys and pays the model usage bills, since those costs sit outside the Chatwoot license. Done when a written note lists the edition, seat count, model provider and the owner of each API account.
Deploy Chatwoot per its self-hosted documentation
Deploy Chatwoot inside your own cloud account or data center using the Chatwoot self-hosted documentation, choosing the Linux VM script, Docker or Kubernetes path that matches your operations model. Check the deployment requirements for Ruby, Node.js, PostgreSQL, Redis and sizing before provisioning, because the requirements page sets the minimum resources. Your team or DeployOpen owns the hosting, backups and upgrades. Done when the instance runs, admin login works and a test conversation reaches the inbox.
Export or pull Intercom content and Help Center articles
Pull Help Center articles and the content Captain will rely on from Intercom. Conversation history is covered in the separate migration pages, so this step only covers the knowledge that answers customers. Export articles in a structured form and record which collections are current. Confirm with Intercom whether any articles are in drafts or restricted collections that the export does not capture. Done when a reviewed inventory lists every article to be reused, with an owner for each.
Build the knowledge source Captain will use
Load the approved articles and answers into Chatwoot's Help Center portal, then connect that content as the knowledge source for Captain. Remove outdated or contradictory answers during review, because an AI agent repeats whatever it is given. Assign a named content owner who approves changes after go-live. Done when a reviewer has checked a sample of common questions against the source and each answer traces to an approved article.
Configure Captain on a pilot inbox
Enable Captain on one low-risk inbox, such as a single website widget or a narrow set of FAQ topics, and define when it hands off to a human agent. Set escalation rules for billing disputes, complaints and regulated topics, and keep the handoff path tested. Use Chatwoot's roles and permissions to limit who can edit Captain settings. Done when test conversations show correct answers, a clean handoff to a named team and a log of each Captain reply.
Run Fin and Captain in parallel
Keep Fin live on its existing inboxes while the pilot runs on Captain, and compare the two on the same question categories. Track resolution rate, handoff rate, customer satisfaction results and any answers a reviewer rejects. Agree the metrics and the review cadence in a written scope, because no duration for this phase is supplied and it depends on your ticket volume. Done when the team has a documented comparison and a signed decision on which topics move.
Cut over and retire Fin by topic
Move topics to Captain in the order the comparison supports, updating widget placement, routing and the handoff rules at each step. Turn off Fin for a topic only after its Captain replacement has passed review, and keep a rollback route in place until the final topic moves. Confirm with your legal and compliance owners how conversation records and model prompts are retained once Intercom is no longer in use. Done when Fin is disabled for every topic, the rollback route is closed and the cutover record is filed.

Running Captain next to the rest of the Intercom stack

Most teams will not switch the Messenger, Help Center and AI agent in one step. These cards cover what can run side by side during a transition, which Chatwoot integrations are documented, and what changes in routing and reporting while two AI agents are live.

Keeping the Intercom Messenger live

The Messenger on your site is the first thing to decide, because two chat widgets on one page will compete for visitors. A common pattern is to keep Intercom's Messenger on pages that are not yet moved, and install the Chatwoot website live chat widget on a limited set of pages, such as one product area, so you can compare answer quality. Confirm with your web team how each widget loads and whether both can share a page without conflicts in cookies or layout.

Help Center content during the overlap

Captain answers from content you give it, so the source articles matter more than the chat tool. While Intercom's Help Center still serves customers, decide whether articles are copied, linked or maintained in both places, and who owns updates when a policy changes. Chatwoot's Help Center portal covers articles, FAQs and guides, so confirm how Captain reads from it before you rely on that content for answers.

Documented Chatwoot integrations

Chatwoot's README lists integrations with Slack, Dialogflow, Shopify, Google Translate and Linear. Use these to connect internal alerts, a bot or flow fallback, order lookups and translation where your current Intercom setup depends on them. Do not assume an Intercom app has an equivalent in Chatwoot. Map each existing Intercom integration to a specific Chatwoot integration or a custom build, and confirm the details with the Chatwoot documentation.

Escalation rules during two AI agents

While Fin and Captain both run, a conversation can be answered by the wrong agent or handed off twice. Define one owner for each channel or page, and write down the handoff path: when Captain passes a conversation to a human, which team receives it, and whether Fin is switched off for that conversation. Test the handoffs with real sample questions before you move traffic.

Routing and business hours

Chatwoot provides auto-assignment by agent availability, business hours, teams and labels. Intercom's routing rules will not carry over automatically, so rebuild the logic in Chatwoot and compare outcomes on the same test conversations. Check how out-of-hours messages are handled by each AI agent, since a bot answering at night can hide a response gap that only appears in reporting later.

Reporting across both systems

Resolution counts from Fin and conversation reports from Chatwoot measure different things, so they will not match during the overlap. Agree on one definition of a resolved conversation and a common CSAT question before the trial starts. Export a sample period from each system and compare volumes by channel, since mismatched definitions are the most common reason a pilot looks better or worse than it is.

What do buyers ask before replacing Fin with Captain?

What drives the total cost of replacing Fin with Captain?

The main cost drivers are the number of agents who need Captain, the hosting footprint, the model access you pay for, and the internal or contracted time to curate knowledge and test answers. On the Chatwoot side, the Community Edition is listed at $0, while the Premium Support plan lists $19 per agent per month and adds Captain AI, so Captain's license cost scales with agent count. Hosting is recurring and grows with conversation volume, since Chatwoot's requirements page recommends 4 CPU cores and 4 GB of RAM for up to 10,000 conversations a day and 8 cores and 8 GB for up to 20,000. Model usage and implementation are separate line items, so ask any vendor or partner for each one in writing.

Can our team deploy and run Chatwoot with Captain ourselves?

Your team can deploy Chatwoot yourselves if you have Linux, PostgreSQL and Redis operations skills. Chatwoot documents a Linux VM installation script, Docker, and Kubernetes with a Helm chart for larger high-availability deployments. The supported runtime is strict: Ruby 3.2 or later, Node.js 20.x, PostgreSQL only, Redis 7.0 or higher, and Ubuntu 20.04 as the only officially supported Linux distribution. Captain depends on the same stack plus connections to model and knowledge sources. If your team lacks these skills, you can hire an implementation partner under a fixed-scope written agreement, which is how DeployOpen works.

What does the work need from our team beyond software?

A self-hosted AI agent needs three things from your side: model access, knowledge curation and testing. Model access means choosing a model provider, creating credentials and deciding where prompts and conversation content are processed. Knowledge curation means organizing help center articles, product documentation and policy text so Captain answers from approved material, and someone must own updates when policies change. Testing means running real-world question sets before go-live and after each change, including escalation paths to human agents. Plan named owners for each of these, because they recur after launch.

How long does a rollout take, and why is no figure given here?

Rollout length depends on the number of channels, the volume and quality of knowledge sources, integrations with your CRM or order systems, and how many test cycles your risk owners require. We do not publish a typical duration because the supplied research does not establish one, and a number without your scope would mislead you. Ask for a timeline in the written scope, and check it against the milestones that matter to you: hosting ready, knowledge loaded, answer testing passed, and cutover with human fallback in place.

Does Captain cover the same resolution logic as Fin?

Captain covers automated answers to common queries, which is what Chatwoot's README describes, but you should not assume feature parity with Fin without testing. Confirm with Chatwoot how Captain handles handoff to humans, resolution tracking, multilingual answers and the actions it can take in your workflows. Then run your own top 50 to 100 conversation types through both the current Fin setup and a Captain pilot, and compare resolution, escalation and wrong-answer rates. Where Captain lacks a behavior you rely on, budget for custom logic or a different agent design.

Where does customer data go when Captain calls an external model?

Your self-hosted Chatwoot keeps conversation records, contacts and knowledge inside your infrastructure, but a hosted large language model receives the text needed to generate each answer. That means data residency depends on the model provider's location, retention terms and contractual commitments. If residency is a hard requirement, ask whether Captain can use a model you host or a provider region you choose, and document the data flow for your privacy and security reviews. Note that DeployOpen prepares systems and evidence for compliance assessments but does not issue audit reports or certifications.

When is the Community Edition with a custom agent enough?

The Community Edition, listed at $0 and free forever, may be enough when you can build and maintain your own agent logic and do not need Captain, SSO, or the per-agent features in the paid tiers. A custom agent means your team or a partner connects a model, writes retrieval and escalation logic, and owns its testing, upgrades and monitoring. Weigh that build and maintenance effort against the $19 per agent per month Premium Support plan, which adds Captain and priority support. Choose the Community Edition only if the engineering ownership is funded and staffed.

Risks to test before switching off Fin

Each risk below changes how a self-hosted AI agent behaves once real customers depend on it. Test all four on your own content and traffic before you turn off Fin, because a gap in any one of them shows up as a worse customer experience rather than a clear error.

Answer quality on your own content

Captain answers from the knowledge sources you connect, so its accuracy depends on how current and well-structured your help articles, policies and product notes are. Build a test set of real customer questions from the last few months, including ones with conflicting or outdated articles. Score each answer for correctness and for whether it cites the right policy before any customer sees it. Confirm with Chatwoot which model providers and knowledge inputs Captain supports in your plan.

Escalation to human agents

An AI agent that fails to hand off cleanly frustrates customers more than one that never tries. Check that low-confidence answers, explicit requests for a person and sensitive topics such as billing disputes or complaints route to a human queue with the full conversation attached. Test handoff during business hours and outside them, when no agent may be online. Confirm how your Chatwoot plan handles assignment rules and business hours.

Model provider lock-in and changing terms

Captain runs on a model provider you choose and pay for directly, so pricing, data handling and model availability can change outside your control. Record which provider and model each Captain feature uses, and what contract terms cover retention and training on your data. Test a second provider on the same question set so you know whether switching is a configuration change or a rewrite. Ask your provider for written notice periods for model deprecations.

Sidekiq and Redis load at volume

AI replies add background jobs and queue pressure on top of normal message traffic, and Sidekiq can use more than 1 GB of memory on a very active server. Use Chatwoot's sizing guidance for your expected daily conversation count as the starting point, then run a load test that simulates peak volume with Captain enabled. Watch queue latency in Redis and memory on the Sidekiq workers, and confirm the swap and Redis versions your deployment needs.

How DeployOpen scopes an AI agent rollout

DeployOpen starts every AI agent rollout with a written scope agreed before any work begins. For a self-hosted Chatwoot setup with Captain, that scope names the environment (cloud account, data center or air-gapped network), the Chatwoot edition and plan that will be licensed, the inboxes and channels Captain will answer on, the knowledge sources it may draw from, and the deliverables and acceptance checks for each. Because most engagements are fixed-scope, the scope document is also where the reader decides what is excluded, such as voice calls, third-party integrations or migration of Intercom conversation history, and where those exclusions are priced separately. Ask for the scope to list the Captain-capable plan and the agent seats it covers, since the Premium Support tier on Chatwoot's self-hosted plans page is the one that adds Captain AI.

Once the scope is signed, the work falls into three parts: implementation, hardening and documentation. Implementation covers the Chatwoot deployment itself, including the PostgreSQL and Redis tiers, the sizing decision based on expected daily conversations, and the connection of the model provider and knowledge sources that Captain uses. Hardening covers access roles, the handling of conversation data sent to the model, logging and backups. Documentation covers the runbook for changing knowledge content, escalation rules to human agents and the evidence a compliance assessor would need. Optional security testing of the AI agent, such as prompt injection and data leakage checks, can be added to the same scope as a separate deliverable. Ongoing support, upgrades and continuous testing are available as retainers, which the reader should price against the scope rather than assume they are included.