Warmly

Authenticated action across distinct classes, with a browsable catalog behind it.

The vendor publishes an integrations directory it describes as 40 or more apps, and names across classes on the product pages: Salesforce and HubSpot as CRMs with native bidirectional sync and OAuth connection, Outreach and Salesloft as sales engagement platforms, Apollo, LinkedIn Ads and Google Ads on advertising, Slack and Microsoft Teams for communication, Google Calendar and Outlook for booking, Marketo for marketing automation, Google Analytics, Bombora for research intent, and Zapier as a general bridge.

Writes go into the customer's own systems rather than staying in Warmly: records are created and updated in Salesforce or HubSpot, contacts are enrolled in sequences, and audiences are pushed to LinkedIn. Webhooks are entitled on every tier for anything not natively covered.

Sourcewarmly.ai/p/pricing and warmly.ai/p/inbound-agent; warmly.ai/p/product/intent-signals/integrationread 2026-09-02

Two agents running a chain end to end and handing work to each other. On site the Inbound Agent identifies a visitor at person level, assembles context from the Context Graph, routes to the owning rep before the first message, engages with AI chat, books directly on the rep's calendar through Google or Outlook respecting availability and round-robin rules, and alerts the team in Slack.

Off site the TAM Agent builds dynamic audiences, tiers accounts by ICP, identifies buying committees, scores intent, enriches, targets LinkedIn ads, orchestrates outbound and runs nurture campaigns. The two are coupled by design: a visitor who does not convert is passed to the TAM Agent for retargeting and nurture, and handed back to the Inbound Agent with updated context on return, which the vendor describes as a continuous loop rather than two products. The chain completes while the visitor is still on the page.

Sourcewarmly.ai/p/inbound-agent and warmly.ai/p/pricingread 2026-09-02

Grounded on the customer's own material on both sides of the conversation. The AI conversation engine is trained on the customer's product, messaging and objection handling, so what the agent says comes from the company's own positioning rather than generic copy, and the vendor states it refuses to answer outside that material rather than inventing.

What it knows about the visitor is assembled live from the Context Graph, a maintained structure the vendor documents in three parts: an Activity Ledger of site visits, emails, call transcripts, CRM activity, chat logs, conversions and ad engagement; an Entities layer of people, companies, deals, buying committees and enrichment data with entity resolution; and an Insights layer carrying ICP tier, intent score and account summary.

That is pulled from CRM, enrichment, website history and intent signals in under 100ms while the visitor is still on the page. The enrichment is largely purchased third party data about the visitor rather than an ingested store of the customer's own documents, so grounding is deep on who someone is and thinner on the customer's wider content.

Sourcewarmly.ai/p/context-graph and warmly.ai/p/inbound-agentread 2026-09-02

Genuine handoff and bounding, with no gate before the agent speaks.

Three shipped mechanisms: the AI qualifies and engages first and then hands off to a rep with full context passed to the human, which is a documented escalation surface Warmly owns; the AI is built to recognize its limits and, on a question outside its training, offers to connect the visitor to a rep or book a meeting rather than answering, which the vendor states plainly it never does inaccurately; and popups carry frequency capping to prevent repeated interruption.

Alert triggers, qualifying criteria and routing rules are all customer-configured, so the conditions under which agents act are set by the team.

Oversight is exercised at configuration time and by taking over mid-conversation, not by approving anything: no review queue, approval step, hold or confidence threshold gates the AI before it speaks to a visitor, sends a follow-up email or adds someone to a retargeting audience, and the top tier is explicitly an autopilot with AI goal-setting, qualification and decision-making. Warmly's writing about agents earning expanded authority through a progressive trust model is thought leadership rather than a shipped feature.

Sourcewarmly.ai/p/inbound-agent and warmly.ai/p/pricingread 2026-09-02

Strong compliance documentation, with no documented access model. Nothing on the trust center or any product page names identity integration or an access model, with no single sign-on, SAML, SCIM or OIDC through the customer's own identity provider, and no role-based access control, roles, granular permissions or least-privilege statement about who inside the buying organization can do what, so a buyer cannot see who on their team can make the agents act.

The compliance side is strong: a SafeBase trust center readable without a call, listing SOC 2, GDPR, CCPA and the EU Data Act, a pentest report, a CAIQ self-assessment, a security whitepaper, app, data and cloud diagrams, BC/DR, SDLC, data security and incident response policies, cyber insurance, a service-level agreement, a published risk profile with a 24-hour recovery objective, and a named responsible disclosure path.

Audit Logging, Access Log Management and Access Monitoring are records of what happened rather than a model of who may act. The attestation is stated as SOC 2 without a type, and the report sits behind an access request with a working path. The product de-anonymizes named individuals, and Warmly places the data controller obligation on the customer.

Sourcesecurity.warmly.airead 2026-09-02

Decision transparency in two places, and no record of what the agent did. Warmly publishes transparent identification, letting a team see exactly how a visitor was matched, which matters because the match is the decision everything else rests on and the vendor states it would rather show unknown than show wrong. The Context Graph surfaces Next Best Action and Reasoning as a named insight alongside ICP tier and intent score, and real-time alerts carry the who, what and why into Slack. Reporting covers chat metrics, campaign attribution and pipeline.

None of that reconstructs agent behavior: no transcript trail, decision log or record of what the AI chat said to a given visitor and on what basis, no history of which retargeting audience someone was added to or why, no configuration audit trail and no retention setting. A customer can see why a person was identified and prioritized, not what was then said to them. The trust center lists Audit Logging among product security features with no published scope, so whether it records agent actions or only account access is not stated.

Sourcewarmly.ai/p/context-graph, warmly.ai/p/inbound-agent and warmly.ai/p/pricingread 2026-09-02

Continuity across visits, held in a layer the agents read back. The AI chat engages with full context from every previous interaction, and the vendor describes the loop explicitly: a visitor who does not convert is captured, nurtured by the TAM Agent, and re-engaged by the Inbound Agent with updated context when they return.

Underneath it the Context Graph keeps an Activity Ledger of what happened, enumerated as site visits, emails, call transcripts, CRM activity, chat logs, conversions and ad engagement, which is a durable record of prior exchanges rather than a snapshot assembled per session. The Context Graph is a shipped product with its own navigation section, a How It Works page, a published component breakdown and a stated sub-100ms query time. It is a shared data store rather than an agent memory layer, and no per-agent state, retention period, inspection or forgetting control is exposed to the customer.

Sourcewarmly.ai/p/inbound-agent and warmly.ai/p/pricingread 2026-09-02

Multi-tenant cloud with no customer control over where anything runs. The trust center names Google Cloud Platform and Firebase as the infrastructure and states nothing about regions; no region selection, EU option, private VPC, single-tenant tier or on-premises route appears on the trust center, the pricing page with its full entitlement lists, or any product page.

The one geography statement the vendor does make is about coverage rather than location: Warmly works globally with visitor identification strongest in the US, UK, Canada, Australia and Western Europe, which describes where the data comes from, not where it is processed. The security whitepaper and the app, data and cloud diagrams listed on the trust center sit behind an access request.

Sourcesecurity.warmly.ai and warmly.ai/p/pricingread 2026-09-02

Two named agents arrive built, and nothing is offered to browse and adopt.

The product is the Inbound Agent for on-site conversion and the TAM Agent for off-site orchestration, each shipping configured rather than authored, and the entry tiers include one AI Studio Agent with unlimited AI Studio Agents on the top tier, which is a surface for the customer to build agents rather than a catalog to select from.

Warmly's Playbooks library under Resources, alongside the blog, academy and webinars, holds authored how-to articles, one of them a step-by-step guide to connecting the MCP to Claude, so a customer reads a play and then builds it rather than adopting a packaged configuration.

Sourcewarmly.ai/p/pricing, warmly.ai/p/inbound-agent and warmly.ai/p/resources/playbooksread 2026-09-02

Wide intake and wide output, both enumerated by the vendor. Signals are published in three tiers: first-party website visits at person and company level, in-app visits, CRM fields and lead scoring; second-party social activity monitoring; and third-party Bombora research intent, competitor keyword monitoring, job changes and champion tracking, new hire detection, job postings, funding announcements, technographic changes, leadership changes and an AI web research signal.

Firing is unusually immediate: identification, context assembly and routing complete while the visitor is still on the page, and alerts fire on configurable combinations of ICP tier, intent score, pages viewed, company size or industry rather than on a schedule. Exit intent triggers popups. Channels out span on-site AI chat, personalized landing pages, smart popups, live human chat, an AI call agent, direct calendar booking, Slack alerts, email follow-up, LinkedIn ads and nurture sequences.

Sourcewarmly.ai/p/pricing and warmly.ai/p/inbound-agentread 2026-09-02

The customer does not choose the model, and no provider is named. The conversation engine is described only as LLM-powered, trained on the customer's product and messaging, and no model selector, per-agent choice, routing policy, bring-your-own-model or bring-your-own-key appears on any product page, the pricing entitlement lists or the trust center.

The Inbound Agent page's Ask AI about Warmly row links out to ChatGPT, Perplexity, Claude and Gemini as a marketing widget, and the MCP server lets a customer's own assistant call Warmly; both are the customer picking the interface that reads the vendor, not the customer choosing what powers the product.

Sourcewarmly.ai/p/inbound-agent and security.warmly.airead 2026-09-02

Two documented routes make Warmly callable from outside. Warmly's own launches page states that its MCP server and REST API are live at opps-api.getwarmly.com, with OAuth-based MCP access from Claude Desktop, Claude Code, Cursor and any MCP-compatible agent, and REST endpoints with API key auth alongside. The vendor publishes working setup instructions, including the claude mcp add command with an organization-ID header, where to find that ID in the product, and the named tools a connected assistant receives: list_warm_visitors, list_warm_accounts, list_third_party_signals and get_credits_remaining.

A later launch extends third-party signals into both the MCP and the API. Webhooks are separately entitled on every pricing tier as SEP and webhook integrations. The MCP tools are read only, so an assistant can retrieve visitors, accounts and signals but cannot act back on the platform. No SDK or client library is published, and access is metered, free with 250 credits a month on every account and paid tiers from $99. A Zapier connector is Zapier's surface rather than Warmly's.

Sourcewarmly.ai/p/resources/launches, warmly.ai/launches/warmly-mcp-and-api-are-live and warmly.ai/p/pricingread 2026-09-02

A real mechanism, aimed at the conversion outcome rather than at the agent. Smart Popups ship with A/B testing built in, so a customer can run one offer against another and read which converts, and that is a controlled comparison against a criterion rather than a dashboard. Around it sit chat metrics, full attribution showing which campaigns drive engagement, pipeline reporting, and an auto-learn and improve loop on the AI chat entitled on the top tier.

The comparison is aimed at the conversion outcome, and nothing tests the agent. No sandbox, dry run, preview of a chat configuration before it goes live, scored evaluation of chat responses, regression check when the training material changes, or comparison of one agent configuration against another is documented, so a change to how the AI talks to buyers is judged only after real visitors have met it.

Sourcewarmly.ai/p/inbound-agent and warmly.ai/p/pricingread 2026-09-02

Everything the agents touch is either the customer's own site or a programmatic interface. Warmly installs as one line of JavaScript on the customer's website, and the on-site surfaces it drives, AI chat, smart popups and personalized landing pages, are rendered by that snippet inside a property the customer owns and has instrumented deliberately, which is the vendor operating its own widget rather than an agent driving software it does not control.

Off site, identification runs through a multi-vendor waterfall querying identity providers in parallel, and every action goes through an API: CRM writes, sequence enrollment, LinkedIn audience sync, Slack alerts and calendar booking through direct Google and Outlook integration. No browser control, headless session, virtual desktop, screen or click automation, or navigation of a third-party console appears anywhere.

Sourcewarmly.ai/p/inbound-agentread 2026-09-02