Agents vs Intercom Fin
Agents is the zerosoft app that replaces Intercom Fin. By our count it covers 98% of the 64 Intercom Fin capabilities we measure: 62 included, 1 partial and 1 missing. The table comes from the same data as the product page and was last checked against the code on October 28, 2026.
Agents isn't usable yet. The table counts what's already built in the code.
At a glance
- Parity with Intercom Fin
- 98%
- Intercom Fin at list price
- $0.99 / resolution
- Agents
- $0 on your server (AGPLv3). On zerosoft Cloud: free plan, then usage. Seats are not billed.
- License of Agents
- AGPLv3, source on GitHub
- Last reviewed
- October 28, 2026
Intercom charges about $0.99 for every conversation Fin resolves, on top of Intercom seats. At 1,000 resolutions a month that is $11,880 a year. Agents on your server costs whatever your AI model uses.
Capability by capability. Our count, from the code.
parity with Intercom Fin
98%of what Intercom Fin does, by our count
- Included
- 62
- Partial
- 1
- Missing
- 1
We measured 64 Intercom Fin capabilities: each one we have scores 1, a partial one 0.5 and a missing one 0, with double weight for the core ones; this is our own assessment, checked against the code on October 28, 2026. See the changelog
- Building the agent100%
- Knowledge100%
- Tools and actions96%
- Testing and QA100%
- Channels90%
- Handoff to a person100%
- Versions and collaboration100%
- Security and admin100%
- Analytics100%
- Answer quality100%
- Import and export100%
- Interface100%
Everything Agents includes69
[x] Included[+] Better[~] Partial
| Capability | Agents | Intercom Fin |
|---|---|---|
| Building the agent | ||
| Free-text instructions | IncludedUp to 12,000 characters | Intercom Fin: Yes |
| Agent templates | Included3 templates (support, SDR that books, lead qualifier), each with its tools preset | Intercom Fin: Yes |
| Name, emoji and welcome message | Included | Intercom Fin: Yes |
| Guardrails | IncludedUp to 30 topics to avoid, a max reply length and spam words. "Check rules" finds rules that collide, with fixed checks and an AI review | Intercom Fin: Yes |
| Deterministic multi-step procedures | IncludedUp to 10 procedures of up to 30 steps: messages, questions with validation (email, phone, number, buttons), forms, if-then branches, tool actions and handoff. They start by keyword or when the agent recognizes the case, and run without improvising | Intercom Fin: Yes |
| Different instructions per channel or audience | IncludedExtra instructions for the Chat widget, the public page and the API | Intercom Fin: Yes |
| Customer memory across conversations | IncludedShort facts per identified visitor (or the same browser on the public page), written by AI when the conversation ends. Visible and deletable in Conversations | Intercom Fin: Yes |
| Identified visitor context (CRM, bookings, issues) | IncludedSigned identity in the Chat widget or a user sent by your server through the API: CRM contact, upcoming Cal bookings and open Track requests | Intercom Fin: Yes |
| Knowledge | ||
| Docs pages as knowledge | IncludedAll public pages or up to 200 picked ones, read live from the same suite with no sync | Intercom Fin: Yes |
| Website content | IncludedUp to 20 URLs plus 3 sitemaps of up to 300 pages each (with a path filter), read in the background within minutes and refreshed daily | Intercom Fin: Yes |
| File uploads (PDF, DOCX, TXT, CSV) | IncludedUp to 25 files of 10 MB. PDFs need selectable text: scanned PDFs (no OCR) and very large or complex ones are not read | Intercom Fin: Yes |
| Question and answer snippets | IncludedUp to 500 pairs with up to 10 variants each, CSV import, plus the canned responses of a Chat widget | Intercom Fin: Yes |
| Import from Notion, Zendesk, Confluence and similar | IncludedZendesk help centers (public, up to 3, refreshed daily). Notion and Confluence through the Docs importer, whose pages work as knowledge | Intercom Fin: Yes |
| Semantic retrieval | IncludedOpt-in per agent: embeddings through the AI Gateway fused with keyword ranking, indexed in the background. Without AI it falls back to keywords | Intercom Fin: Yes |
| Tools and actions | ||
| CRM: find, create or update contacts, log notes | BetterNative to the suite CRM, no connector to set up | Intercom Fin: Yes |
| Book meetings with real availability | IncludedUses a Cal event type, looking up to 30 days ahead | Intercom Fin: Yes |
| Create a ticket or issue | IncludedCreates an issue in a Track project | Intercom Fin: Yes |
| Run an automation or a scraping robot | BetterUp to 20 Flows and 20 Robots per agent, from the same suite | Intercom Fin: Yes |
| HTTP requests to your own APIs | IncludedUp to 20 allowed domains and 10 secret headers | Intercom Fin: Yes |
| Human approval before write actions | IncludedPer tool, for 6 write tools (HTTP only when not GET) | Intercom Fin: Yes |
| MCP server connections | IncludedUp to 5 streamable HTTP servers with an encrypted auth header that is hidden from replies and logs. Tools that change data wait for an admin outside the playground | Intercom Fin: Yes |
| Buttons and forms inside the chat | IncludedQuick-reply buttons and up to 5 forms (8 fields) in the Chat widget, the public page and the playground, validated on the server and saved in the CRM. Text channels get numbered options | Intercom Fin: Yes |
| Stripe and Shopify actions | PartialRead only, with your own keys: Stripe billing of the identified visitor (subscriptions, invoices) and Shopify order status and product search. Refunds, cancellations and other writes are not available | Intercom Fin: Yes |
| Custom code steps | IncludedUp to 10 JavaScript functions the agent or a procedure can call, run in an isolated Vercel Sandbox machine with no network (self-hosting needs Vercel Sandbox credentials) | Intercom Fin: Yes |
| Testing and QA | ||
| Playground | IncludedWrite tools are simulated, with the input and output of each call shown | Intercom Fin: Yes |
| Saved tests graded by an AI judge | IncludedUp to 50 tests, saved from any playground reply | Intercom Fin: Yes |
| Multi-turn simulations with an AI customer | IncludedAn AI customer with a persona and a goal, up to 8 turns, graded as a whole by an AI judge | Intercom Fin: Yes |
| Batch replay of real conversations | IncludedThe last 10 real conversations replayed against the draft (simulated), side by side, with an AI verdict | Intercom Fin: Yes |
| Channels | ||
| Website chat widget | IncludedAnswers as the bot of a Chat widget, with handoff to the inbox | Intercom Fin: Yes |
| Public agent page | IncludedHosted at /a/<handle> | Intercom Fin: Yes |
| API access | IncludedPOST /api/v1/agents/:id/messages with an API key | Intercom Fin: Yes |
| Agent as a step in any automation | BetterA Flow node, usable outside the chat too | Intercom Fin: Yes |
| WhatsApp, Instagram, Messenger and Slack | IncludedWhatsApp through Lines (new one-to-one chats or chat by chat), and Slack, Messenger and Instagram with your own Slack or Meta app (tokens encrypted, signed webhooks) | Intercom Fin: Yes |
| Email channel | IncludedAnswers emails sent to the address of its Chat widgets, in the same thread (needs incoming email set up in Chat) | Intercom Fin: Yes |
| Voice and phone | IncludedAnswers calls turn by turn (speech recognition and spoken reply) on a number of the instance’s Twilio account that the instance admin assigns to the workspace, and transfers to a phone number on handoff. Not a real-time voice model | Intercom Fin: Yes |
| Handoff to a person | ||
| Handoff on request, when unsure or by keyword | IncludedUp to 30 keywords | Intercom Fin: Yes |
| Custom handoff message | Included | Intercom Fin: Yes |
| Team replies from the agent or the Chat inbox | Included | Intercom Fin: Yes |
| Agent stops replying after handoff | IncludedOn every channel | Intercom Fin: Yes |
| Handoff when the AI model fails | BetterOn the public page the visitor gets the handoff message and the team a notification | Intercom Fin: No |
| Handoff reasons in reports | IncludedKeyword, model failure or the reason the agent gave, counted in Analytics | Intercom Fin: Yes |
| Versions and collaboration | ||
| Draft and published versions with rollback | IncludedAfter the first publish, saving creates a draft that the playground and tests use. Any version can go back live. Files, Q&A pairs and MCP servers apply right away | Intercom Fin: Yes |
| A/B tests of agent settings | IncludedA saved version answers a share of new conversations (5 to 95%), compared on resolution, handoffs, thumbs and CSAT; put the winner live in one click | Intercom Fin: Yes |
| Log of who changed what | IncludedWho saved, published, restored or changed knowledge and servers, in the Versions tab | Intercom Fin: Yes |
| Read-only roles and permissions per agent | IncludedPer agent: members can view, edit or not see it (default and per person), on top of the per-app access of the workspace. Editors only turn on tools of apps they have, and the playground reads customer data only for them. Secrets, connections and approvals stay with admins | Intercom Fin: Yes |
| Security and admin | ||
| Encrypted secrets the model never sees | IncludedSecret headers, MCP tokens and the Stripe, Shopify, Slack and Meta keys are encrypted, shown only by their last 4 characters and covered in replies, logs and exports. Only admins see or change them | Intercom Fin: Yes |
| Domain allowlist and blocked private network calls | Included | Intercom Fin: Yes |
| Message limits per conversation and per workspace | IncludedDefaults: 20 per conversation every 10 minutes, 300 per workspace per hour | Intercom Fin: Yes |
| Personal data redaction in transcripts | IncludedEmails, phones, cards and ID numbers by pattern, plus names and street addresses found by AI (optional), an hour after the conversation goes quiet and again if it resumes | Intercom Fin: Yes |
| Spam filtering and visitor blocking | IncludedSpam words, blocked countries on the public page and visitors blocked from their conversation (public page and Chat) | Intercom Fin: Yes |
| Analytics | ||
| Resolution, handoffs, conversations, response time | IncludedLast 7, 30 or 90 days | Intercom Fin: Yes |
| Errors per tool and runs that hit the step limit | BetterWith the last error message of each tool | Intercom Fin: No |
| Conversation topics | IncludedAI topic per scored conversation (when AI scoring is on), with keyword groups otherwise | Intercom Fin: Yes |
| Thumbs up and down on replies | Included | Intercom Fin: Yes |
| Email alerts on spikes | IncludedHourly check of thumbs down, tool errors and step limits against the prior week | Intercom Fin: Yes |
| AI-scored CSAT and sentiment per conversation | IncludedCSAT 1 to 5, sentiment, resolution and topic, half an hour after the conversation goes quiet. Turned on per agent, up to 300 a day | Intercom Fin: Yes |
| Breakdown by channel | IncludedConversations, resolution and handoffs per channel, Flow apart from the API | Intercom Fin: Yes |
| Models | ||
| Choose the AI model | BetterAny provider/model from the AI Gateway | Intercom Fin: No |
| Temperature setting | BetterFrom 0 to 1 | Intercom Fin: No |
| Clear model errors for the owner | BetterSays where to change the model; visitors see a neutral message | Intercom Fin: No |
| Answer quality | ||
| Answers grounded in your content | IncludedSearches the knowledge base before answering and hands off when unsure | Intercom Fin: Yes |
| Source citations and links in replies | IncludedNumbered sources at the end of the reply, with a link for help center pages and web pages | Intercom Fin: Yes |
| Conversation quality review and scorecards | IncludedUp to 10 scorecard criteria graded by AI, a "To review" queue and reviewed marks with notes | Intercom Fin: Yes |
| Unanswered questions turned into content suggestions | IncludedQuestions the knowledge could not answer in the last 30 days, grouped, with "Add answer" to turn them into a Q&A pair | Intercom Fin: Yes |
| Import and export | ||
| Export conversations | IncludedCSV with one row per message and the score, for the last 7, 30 or 90 days, plus the API | Intercom Fin: Yes |
| Interface | ||
| From template to editor in one click | IncludedPublic Docs pages are already set as knowledge | Intercom Fin: Yes |
| Keyboard tabs and a single save with Cmd+S | Included | Intercom Fin: No |
| Find agents from the command menu | Included | Intercom Fin: Yes |
| Editor works on mobile and in dark mode | IncludedNo horizontal scroll at 390 px | Intercom Fin: Yes |
Where we go further
- CRM: find, create or update contacts, log notesNative to the suite CRM, no connector to set up
- Run an automation or a scraping robotUp to 20 Flows and 20 Robots per agent, from the same suite
- Agent as a step in any automationA Flow node, usable outside the chat too
- Handoff when the AI model failsOn the public page the visitor gets the handoff message and the team a notification
- Errors per tool and runs that hit the step limitWith the last error message of each tool
- Choose the AI modelAny provider/model from the AI Gateway
- Temperature settingFrom 0 to 1
- Clear model errors for the ownerSays where to change the model; visitors see a neutral message
Still missing (1)
Intercom Fin has these and we don't yet. If you need one, vote for it on the public board.
- Mobile SDK (iOS, Android)No native SDK. A mobile app can use the API (POST /api/v1/agents/:id/messages), which returns the buttons and forms to drawVote on the board
Questions.
Which AI model does it use?
Any model on the AI Gateway, in provider/model format (for example, anthropic/claude-sonnet-5-5). On your server you pay usage directly with your key; on zerosoft Cloud it is billed by tokens.
Can it make up answers?
It answers from your knowledge and hands off to a person when it is unsure. Every reply is logged with the tools it used, so you can review it.
Where does it talk to my customers?
In the Chat widget on your site, on its own public page, by email (with Chat's inbound email set up) and through API v1 (agents:run scope). Also on WhatsApp through Lines, on Slack, Messenger and Instagram with your own app, and by phone on a Twilio number assigned by the instance admin.
Can it change my data?
Only through the tools you turn on. Suite tools act inside your workspace; outside it, it only calls the domains you allow and your MCP servers. Tools that write can require human approval. To book a meeting it needs a slot it offered itself and the visitor’s confirmation.
Can I bring my Fin setup?
There is no importer. Paste your instructions as the brief and add your Docs pages and the URLs of your web pages as knowledge.
Guides and posts
- Migrate from Intercom to Chat: bring your conversationsHow to move from Intercom to Chat: pull the JSON from Intercom's API, what imports (people, messages, status), what doesn't, what to check first.
- How to test an AI agent before it talks to your customersA four-step method to test a support agent: saved tests graded by an AI judge, simulated conversations, replays of real chats, and versions with a draft.
- How to write the brief and the knowledge for a support agentA brief template for an AI agent, how to write Docs pages the search will actually find, and a test method with real questions before you publish.
- When an AI agent should hand the conversation to a personFive situations where a support or sales agent has to hand off, how to set up each one in Agents, and what to review every week to tune the rules.