Circleback

The pricing page names "automations with all integrations, including Notion, HubSpot, Salesforce, Attio, webhooks, and many more apps" from the Pro tier, with limited automations on Free. Circleback's own writing adds Slack, Linear, Google Docs, Zoho and monday.com to the native list, with Zapier and Make covering the long tail. Calendar and email connection across Google, Outlook and Gmail is the ingestion side.

The agent acts through these systems rather than beside them, which is the point of the product. Meeting outcomes do not stop at a summary: tasks are created in Linear, boards updated in monday.com, Notion databases populated, and deals and contacts updated in HubSpot and Salesforce. Writing a record into a customer's CRM is a state change in a system Circleback does not own.

Action items are assigned to the relevant participants and then routed into the destination system.

No integrations directory is published, so the estate is named rather than enumerated, and the vendor claims no integration count.

Sourcecircleback.ai/pricing tier feature lists with Circleback's own integration listread 2026-09-12

A customer facing automation builder runs the work, not a fixed pipeline. Circleback describes a "native automation builder with no code or third-party tools," and the pricing page tiers it: limited automations on Free, automations with all integrations on Pro, team automations and default and enforceable settings higher up. A customer defines which insights to extract and where each should land, so the sequence is composed by the buyer.

The sequence holds state across a real gap and fans out. A meeting ends; the transcript is produced; insights and action items are extracted; each action item is assigned to the relevant participant rather than listed for someone to redistribute; and each is routed to a different destination depending on what it is (a Linear task, a monday.com board, a Notion row, a HubSpot or Salesforce record), while a summary email goes out unprompted. Fan out to different systems on the content of the item is branching, not a linear pipeline. Naming an owner is what turns a summary into a handoff.

There is no multi agent handoff and no named runtime; what exists is a customer configurable routing engine with conditional destinations. Whether conditional logic is customer definable or destination only is not stated.

Sourcecircleback.ai/pricing with Circleback's own automation descriptionread 2026-09-12

The corpus is a maintained archive, not per run assembly. Every captured conversation accumulates into a searchable store, and the pricing page tiers its lifetime as a product feature: "30-day meeting and recording history" on Free, "unlimited meeting history" on Pro, two years of recording storage on Business and unlimited on Enterprise. A vendor that sells history length by tier is selling a maintained structure.

Retrieval is a named surface: Ask Circleback on every tier, with AI search across past conversations returning clips, summaries and decisions. Circleback describes the assistant as carrying email context as well as meeting context, and the MCP connector as spanning meetings, calendar and email, so the grounded corpus takes in the calendar and the mailbox around each conversation. AI search across shared meetings is a separate team tier entitlement, so the retrieval scope is itself an administered boundary.

The corpus is well conditioned for retrieval: speaker labeled transcripts with automatic participant identification online and in person, support for more than 100 languages, and automatic tagging. Speaker attribution is what lets a retrieved passage answer who said it. Grounding shows in the output: action items carry owners and deadlines pulled from the conversation, and custom AI insights let a customer define what to extract.

Citation back to a timestamp, chunking, and freshness or confidence signals on a search result are not documented.

Sourcecircleback.ai/pricing tier feature lists with Circleback's own assistant and MCP descriptionsread 2026-09-12

Control comes from a design time bound plus a human destination for the output. Automations are configured by the customer before they run, so what the system may extract and where it may write is decided in advance rather than by the agent at runtime. The Business tier adds "default and enforceable settings" and "admin and member roles," putting that configuration under administrative control. And action items are assigned to the relevant participant, so the work lands in front of a person who must act on it.

Nothing gates the agent's own action. No approval step, review queue, confidence threshold or confirmation is documented before an automation writes to a CRM record, creates a Linear task or populates a Notion database. The write happens because the meeting ended and the rule matched. For a product whose selling point is that meeting outcomes reach systems of record automatically, that is the design.

The Free tier lists "view and draft emails" (draft, not send), but nothing states that sending is withheld by design. Assignment is not a gate either: naming an owner routes the work, it does not pause the agent.

Sourcecircleback.ai/pricing tier feature listsread 2026-09-12

In Circleback's own security article, "Circleback is SOC 2 Type II and EU-U.S. Data Privacy Framework-certified," with HIPAA and an external penetration test whose report is published on the security portal. On access, the public pricing page lists single sign on (SAML or OIDC) and SCIM provisioning at Enterprise, admin and member roles and default and enforceable settings at Business, and admin controlled data retention.

The supporting estate is deeper than the company's size suggests: encryption at rest and in transit, an explicit "we do not use customer data to train models," GDPR processor status under a published DPA, a published BAA, the UK Extension to the EU-U.S. DPF for UK transfers with ICO cooperation named for unresolved complaints, a status page, and a trust portal at security.circleback.ai. The certification named is the current framework, not the defunct Privacy Shield.

HIPAA is Enterprise only. The pricing page lists HIPAA compliance only under Enterprise, and the security article says "HIPAA is included with Enterprise." SSO, SCIM and HIPAA all sit on Enterprise at $35 per user per month, so a Free or Pro buyer gets none of the three, which is ordinary enterprise packaging.

Sourcesupport.circleback.ai security article, updated 2026-09-12, with circleback.ai/pricing tier featuresread 2026-09-12

Every output is inspectable and traceable to its origin: speaker labeled transcripts, structured summaries, and action items carrying owners and deadlines pulled from the conversation, with AI search across every past meeting to retrieve the clip or decision behind a claim. Assigned items are then visible in the destination system (the Linear task, the Notion row, the CRM record), so the result of an automation is observable even though the automation itself is not.

The Business tier lists a usage dashboard, which is operational visibility across a workspace, but it reports usage rather than agent activity. A status page at status.circleback.ai covers service availability.

No automation run log, execution history, record of which automations fired on which meeting and what each wrote, failure or retry surface, or export of agent activity is published. For a product that writes into CRM records and task systems automatically, nothing published shows which rows Circleback created and when.

Sourcecircleback.ai/pricing with support.circleback.ai security articleread 2026-09-12

The store is the archive of captured meetings, which is ingestion rather than state the agent writes as a consequence of its runs.

Circleback derives action items, owners and deadlines from a conversation, but those are pushed out to Linear, Notion, HubSpot and Salesforce as the product's whole purpose, so the derived state lives in the customer's destination systems rather than persisting in Circleback to shape a later run.

Nothing describes a prior decision being reapplied to a later meeting. API, MCP and CLI access exist, but no page states that an outside caller can change the retained record.

What it does have is an archive that is scoped, inspectable and administered. It is scoped to a workspace, with team level boundaries on shared meetings and AI search across them sold as a tier entitlement. It is inspectable through Ask Circleback and AI search across every past conversation. And its lifetime is administered: "workspace admins can also control how long meeting data is kept," with custom data retention from the Business tier and default lifetimes stated per tier. Whether retention is set per workspace only or per meeting is not stated.

Sourcesupport.circleback.ai security article and circleback.ai/pricingread 2026-09-12

Data rests in the United States and there is no choice. Circleback's security article describes "transfers of personal data from the EEA or the United Kingdom to Circleback in the United States," relying on its EU-U.S. Data Privacy Framework certification including the UK Extension, with ICO cooperation for unresolved UK complaints.

No region list, region picker, single tenant tier, private cloud, on premises or self hosting option is published. The Enterprise tier at $35 per user per month buys SSO, SCIM, invoice billing, unlimited recording storage, priority support and HIPAA, not a deployment choice.

Bot free capture is a capture modality, not a residency option: the desktop app records locally and the data still lands in Circleback's cloud.

Retention is not residency. Circleback publishes real retention control ("workspace admins can also control how long meeting data is kept"), with custom data retention from the Business tier and recording storage of 30 days on Free, one year on Pro and two years on Business. However long data is kept, it is kept in the United States for everyone.

Sourcesupport.circleback.ai security article and circleback.ai/pricingread 2026-09-12

Working defaults are not a pack. Automatic action item identification and assignment, post meeting summary emails and native connectors are the product doing its job on first run, not units a buyer selects among.

There is nothing to select. No agent gallery, template library, blueprint catalog, industry or role pack, marketplace or contribution path is published. The closest surface is "custom AI insights" on the Pro tier, and custom is the point: the customer authors what to extract, which is an authoring capability rather than a catalog of prebuilt units.

The tiers are a pricing structure, and "limited automations" on Free rising to "automations with all integrations" on Pro is a metering boundary, not a set of adoptable agents. There is no published library of automation templates or recipes, and no named prebuilt insight extractors.

Sourcecircleback.ai/pricing tier feature listsread 2026-09-12

Event triggers fire without invocation. Calendar and email connection across Google, Outlook and Gmail lets Circleback detect scheduled meetings and join them automatically, so the run starts from an event in the customer's calendar. Automations fire on meeting completion, and post meeting summary emails send without prompting.

Channel coverage is wide and named on the vendor's own pages: Zoom, Google Meet and Microsoft Teams for video; Slack huddles, which many meeting bots cannot reach; and in person conversations through desktop, mobile and Apple Watch apps. Delivery runs outward through summary email, Slack and the destination systems themselves.

There are two capture modes, and the choice is the buyer's: "a bot or the bot-free desktop app for online meetings," listed on the Free tier. Bot free local recording is how the product enters meetings where a visible bot would change the conversation.

In person capture with automatic participant identification is uncommon among meeting bots: one tap in person recording on iOS and Android, with speakers identified online and in person and more than 100 languages supported. Events, scheduled meetings and multiple channels are all documented and available on the free tier. No page says whether a customer can define trigger conditions beyond meeting completion.

Sourcecircleback.ai/pricing Free tier feature list with Circleback's own capture and mobile descriptionsread 2026-09-12

Nobody outside Circleback chooses the model. No model picker, provider selection, per automation model assignment, reasoning effort control or bring your own key path is published. The tier ladder lists entitlements feature by feature, down to Apple Watch support and the number of teams, and names no model capability at any level, including Enterprise.

No provider is named at all, so the stack is undisclosed rather than visibly single provider, and the security article and pricing page list no subprocessors.

The one model adjacent statement is about data handling: "we do not use customer data to train models" is a commitment about the customer's content and says nothing about which model runs.

Circleback ships an MCP server, but its direction is inbound: an outside assistant such as Claude or ChatGPT calls Circleback for context, which is not evidence that Circleback routes across models.

Sourcesupport.circleback.ai security article with circleback.ai/pricing tier listsread 2026-09-12

Circleback's pricing page lists "API, MCP, and CLI access" as a feature of the Free tier: not gated, not Enterprise only, not roadmap. Three external surfaces are named on the vendor's own commercial page at zero cost, an unusually low barrier for developers. Webhooks appear alongside the integration list from the Pro tier upward.

The MCP server is inbound and spans more than transcripts. Circleback describes the MCP connector as spanning meetings, calendar and email, reachable from Claude, ChatGPT or Cursor, so an outside assistant can query the meeting corpus, the calendar context and the email context the product holds. A CLI alongside it makes the platform scriptable from a terminal without building against the API directly.

The MCP server is the customer's assistant calling Circleback; the outbound connectors to HubSpot, Salesforce, Linear, Notion, Attio and monday.com run the other way.

No endpoint reference, MCP tool list, SDK, OpenAPI specification or developer host is published, so the surfaces are named on a pricing page and in vendor prose rather than documented for a builder.

Sourcecircleback.ai/pricing Free tier feature list, with Circleback's own MCP descriptionread 2026-09-12

Nothing puts a change under test. No sandbox or test workspace, dry run of an automation before it goes live, preview of what a rule would have written, versioning or rollback of an automation, holdout or comparison, evaluator, or readable score of extraction quality is published. A customer builds an automation and it fires on the next matching meeting; nothing exercises the rule against a past conversation first.

The gap is sharper than it looks. Extraction accuracy is the product's core claim, action items with the right owner and deadline, and the writes land in someone else's system of record. A buyer has no published way to measure how often an extraction is wrong, and no audit trail to find out afterwards, so automation quality is unverifiable from outside.

The Business tier's usage dashboard measures consumption, not the quality of the agent's output.

Sourcecircleback.ai/pricing tier feature listsread 2026-09-12

Every path in and out is a protocol or a platform integration. Capture runs through the Zoom, Google Meet, Microsoft Teams and Slack meeting platforms; outbound action runs through native connectors to HubSpot, Salesforce, Linear, Notion, Attio and monday.com plus Zapier, Make and webhooks; and inbound access runs through an API, an MCP server and a CLI. A third party redesigning its interface does not touch any of it.

The one adjacent capability is audio capture, not interface control. Circleback offers "a bot or the bot-free desktop app for online meetings," and the desktop app records system audio locally without joining the meeting as a visible participant. Recording what comes out of the speakers is passive capture; the agent is not clicking, navigating, filling a field or driving a window. The same applies to the mobile and Apple Watch apps, which capture in person audio on a tap.

No hosted or local browser, desktop session control, remote computer control, RPA, recorder that replays UI actions or browser extension that acts on a page appears. Where the product reaches a system without an API, it uses Zapier, Make or a webhook rather than driving the interface.

Sourcecircleback.ai/pricing with Circleback's own desktop and mobile capture descriptionsread 2026-09-12