Build AI Agents in HubSpot: The Customer-Built Layer of Breeze (2026)

You can now build AI agents in HubSpot without writing code, using Breeze Studio on a Pro or Enterprise plan. The stack has three no-code pieces: custom assistants/agents in Breeze Studio, AI workflow actions (including the new Run Agent action in beta), and Smart Properties powered by the Data Agent.

For developers, HubSpot exposes a remote MCP server at mcp.hubspot.com (GA April 13, 2026), a local Developer MCP server (GA February 19, 2026), Agent Tools, UI Extensions, and custom-code workflow actions.

Knowledge Vaults are still beta and have no public ingestion API. Credit allocations are 500 / 3,000 / 5,000 per month on most hubs, doubled to 5,000 / 10,000 on Data Hub and Customer Platform.

Most RevOps teams still think Breeze is the five HubSpot-built agents.

Customer Agent. Prospecting Agent. Content Agent. Data Agent. Knowledge Base Agent.

That mental model was correct in 2025. It is wrong now.

Between January and April 2026, HubSpot shipped a customer-built layer on top of Breeze that almost no one is talking about. You can now build AI agents in HubSpot — your own agents, with your own knowledge, your own tools, and your own triggers — without writing code.

If you want code, the surface goes further. There is a remote MCP server. A developer MCP server. A framework for building tools that Breeze Agents can call. React-based UI extensions inside CRM records.

This piece is a working read on both layers — the no-code stack a RevOps lead can use this week, and the developer surface your engineering team should be tracking. Where the layers don't work, you'll see that too.

Let's get into it.

The shift no one's framing correctly

HubSpot stopped adding AI features in 2026. It started building an AI platform.

Look at what HubSpot actually released this year:

• January 12 — Breeze Studio custom agents migrated to GPT-5 by default. Audit Cards rolled out, showing every CRM property an agent touched, before and after.
• February 19 — the local Developer MCP Server went generally available, plugging HubSpot's developer context into Cursor, Claude Code, VS Code, and other agentic IDEs.
• April 13 — the remote HubSpot MCP server at mcp.hubspot.com graduated from beta to GA, with full read/write across CRM objects.
• April 14 — Customer Agent flipped to $0.50 per resolved conversation. Prospecting Agent flipped to $1 per recommended lead. Outcome pricing, not consumption.

Read those four moves together. They tell one story: the platform thesis.

The five stock agents are the shrink-wrapped demo. The customer-built layer is the platform.

That distinction matters. Because if you sell into Breeze the way the five stock agents are marketed — "pick an agent, turn it on" — you'll under-spec what your CRM can do by an order of magnitude.

Bottom line. Stock Breeze = the demo. Breeze Studio + AI workflow actions + Smart Properties + MCP = the platform. Most RevOps teams are buying the demo and leaving the platform on the table.

Need a HubSpot partner to map out where AI agents actually fit your revenue motion? Talk to Superwork about a Breeze architecture session before you turn anything on.

Part 1 — The No-Code Track (for RevOps and admins)

This is the layer a Pro or Enterprise admin can ship without ever opening a code editor.

It has three pieces. Most teams use one. The advantage compounds when you use all three.

Piece 1: Breeze Studio

Breeze Studio is HubSpot's no-code workspace for building, customizing, and managing AI assistants and agents inside HubSpot. You find it under Breeze → Breeze Studio in the top nav.

It is the same surface where HubSpot's stock Breeze Agents live. Studio is where you customize them, clone them, build new assistants and agents from scratch, and install agents from the Breeze Marketplace.

HubSpot's own framing is clean: "Breeze marketplace is where you discover, browse, and install HubSpot-built agents and custom assistants… Breeze studio is where you manage and customize Breeze Agents and custom assistants to work for your specific business needs."

As of May 2026, Studio is still labelled BETA in the HubSpot Knowledge Base. Treat it as production-capable. Expect UI changes month to month.

What tier do you need to build AI agents in HubSpot?

Breeze Studio is available on Starter, Professional, and Enterprise plans across Marketing, Sales, Service, Content, and Data Hubs.

The gating that actually matters:

• Starter — basic custom assistant creation. Most agent triggers, most tools, and Marketplace install are locked.
• Professional — full Studio. Automation triggers on agents. Marketplace install.
• Enterprise — higher credit allocations, advanced MCP/tool configurations, and access to Breeze Marketplace partner agents.
• Free CRM / Free tools — no Studio access. Only the basic conversational Breeze Assistant.

Per the HubSpot KB: "A Professional or Enterprise subscription is required to configure automation."

Note. If a prospect is on Starter and wants triggered agents, they're not a Breeze Studio buyer — they're a Pro upgrade conversation first.

Assistant vs. agent: what's the difference?

Both live in Studio. They are not the same thing.

A Breeze Assistant is conversational. You summon it inside Breeze Assistant to answer questions, generate content, or run a process you defined. One-off, interactive.

A Breeze Agent follows defined steps with inputs, tools, and knowledge to complete structured tasks. It can run on a trigger.

HubSpot's own phrasing: "Assistants prioritize speed and guidance, whereas agents may take longer due to task complexity."

The side-by-side:

The decision rule: interactive and ad hoc → assistant. Triggered and structured → agent.

How do you actually build one in Breeze Studio?

Open Studio. Click Create assistant or Create agent. Then five fields matter.

Inputs. The variables the user fills in each run — or that the workflow trigger passes in. Text, dropdown, record reference.

Instructions. Free-text system prompt. The agent's role, constraints, tone, decline rules, output format. For cloned stock agents, this is where you stack "extra instructions" on top of HubSpot's locked defaults.

Tools. The agent's capabilities. Get data tools (search CRM, browse the web). Generate tools (summarize record, draft content, infer ICP). Take action tools (create/update records, create tasks, send emails). Every "Take action" tool ships with a Review before running toggle that forces human approval — on by default.

Knowledge. One or more Knowledge Vaults attached to the agent.

Triggers. On Pro and Enterprise. On a schedule, on record creation, on enrollment in a list. Daily cap is 1,000 agent runs per day, resetting at 00:00 UTC. Bulk enrollment of existing records is not supported at publish.

You test in the right-hand preview pane. You publish when it's ready. You revert through change history if something breaks.

Pro Tip. Always keep Review before running on for any Take action tool during the first month in production. Switch it off only after you've audited 50+ runs and the agent has not produced a single incorrect write.

Can you edit HubSpot's stock Breeze Agents?

Partly. You can layer on top. You can't remove the defaults.

Per the HubSpot KB: "HubSpot's pre-built agents have default inputs, instructions, and tools that can't be edited, deleted, or rearranged."

When you open the Customer, Prospecting, Content, Data, or Knowledge Base Agent in Studio, you can add extra instructions, extra inputs, extra tools, extra knowledge, and configure triggers. You cannot remove HubSpot's defaults.

If you want to go further — different default tone, different default knowledge, a Norwegian or German-first system prompt — you Clone the agent. The clone is a fully editable custom agent.

This is the move when a Nordic mid-market client wants a Prospecting Agent that knows their DACH playbook, or a Customer Agent that defaults to Norwegian for Tier-1 support.

Piece 2: AI workflow actions

AI workflow actions are deterministic, single-step LLM calls you drop into any HubSpot workflow. They are the layer most teams skip — and it's where the highest leverage usually sits.

You find them under Automation → Workflows → AI.

The current lineup:

• Ask OpenAI Assistant and Ask Anthropic — first-party HubSpot actions that send a prompt plus record data to the chosen model and store the response as an action output.
• Use a custom LLM (beta) — bring your own API key for OpenAI, Anthropic, Cohere, xAI, or Gemini. You pick the model, set temperature, write the prompt, reuse the output downstream.
• Summarize record (beta) — one-click record summary. Cheaper than a generic LLM call.
• Infer company value proposition and ICP (beta) — analyzes a company record, returns positioning text.
• Data Agent: Custom prompt / Research / Fill Smart Property (beta) — triggers Data Agent operations from any workflow trigger.
• Run Agent (beta, 500 executions/day cap) — the action that changes the game. Triggers any Breeze agent — stock or custom — from inside a workflow. Configurable input payload. Output as text or as structured fields you define (text, number, boolean, date, datetime, enumeration, phone).

The Run Agent action converts Breeze from "user-invoked" to "system-invoked." It is what turns Studio agents into proper automation primitives.

Pro Tip. Run Agent is free during beta. Per the HubSpot KB, it will switch to credit-billed at GA. If you're architecting a Run Agent–heavy automation, model the post-GA cost into your client proposal now.

When to use a workflow action vs. a full agent

Workflow actions are deterministic. Single step. Predictable cost.

Agents are non-deterministic. They plan tool use across multiple steps. Less predictable.

If the logic is fixed, it's an action job. If the logic is "figure it out," it's an agent job.

Piece 3: Smart Properties

A Smart Property is an AI-researched CRM field powered by the HubSpot Data Agent. You define the field like any custom property — but instead of being filled by users or imports, the value is populated by an AI prompt.

Setup: Data Management → Data Agent → Create smart property. Pick the object. Pick the field type. Write the prompt. Pick the data source — web research, company website, property data, or call transcripts.

You can insert property tokens into the prompt to reference other fields on the record.

Where Smart Properties earn their keep on a Nordic mid-market client

Four prompts that consistently pay for themselves:

• Tech stack detected — "List the marketing-automation, CRM, and analytics tools listed on this company's website."
• Hiring intent signal — "How many open engineering roles are listed on this company's careers page?"
• Funding stage — "Has this company raised a Series B or later in the last 12 months? Return Yes/No."
• Nordic ICP qualifier — "Is this company headquartered in Norway, Sweden, Denmark or Finland and does it have 50+ employees? Return Yes/No plus headcount estimate."

Cost is 10 credits per Smart Property fill, per record, charged whether or not a value is returned.

For 5,000 companies that's 50,000 credits. Burn the whole month's allocation in a single bulk run if you're not careful.

Pro Tip. Never run Smart Properties blindly across the database. Run them selectively on segments — your active sales pipeline, your top 100 ICP accounts, your at-risk renewals. Treat them like paid enrichment, not free enrichment.

What Smart Properties won't do

No validation rules. No Sensitive Data flagging. LinkedIn-based prompts ("search LinkedIn company profile…") are filtered out.

If you need structured truth (validated email, normalized phone, deduped company), Smart Properties is the wrong tool. That's a Data Hub job.

Knowledge Vaults — the part that decides whether your agent is useful

A HubSpot Knowledge Vault is a managed context store you attach to a Breeze agent so it can answer with grounded, citable information from your own documents and CRM data.

Per the HubSpot KB (still labelled BETA in May 2026), a vault can contain:

• Files — .pdf, .md, .html, .pptx, .txt UTF-8, .docx up to 50 MB. .jpeg/.png/.webp up to 10 MB.
• HubSpot content — knowledge base articles, landing pages, website pages, blog posts, call transcripts.
• CRM objects and segments — including custom objects via segments (added in early 2026).
• Notes on CRM records.

Limits worth knowing: up to 50 vaults per account. A vault can't be saved without at least one file or object attached. A default vault is auto-created from your AI settings.

When an agent cites a vault document in a response, the citation surfaces the document name and — where available — page or section.

The Nordic mid-market pattern that works:

One vault per business unit or brand. One vault for the ICP definition and target account list. One vault for the master answer library (RFP responses, security FAQs). One vault per customer-segment for CSM agents.

Bottom line. Don't build a single "everything" vault. Citations get noisy, the agent pulls irrelevant context, and answer quality drops fast. One vault, one job.

The Breeze Marketplace

The Breeze Marketplace is HubSpot's catalogue of installable AI agents, assistants, connector apps, and Agent Tools — both HubSpot-built and partner-built.

You find it at Breeze → Marketplace or ecosystem.hubspot.com/marketplace/breeze-agents.

Click Add on a listing → the agent appears in Studio → Configure to layer your extra instructions, tools, knowledge, and triggers → Publish.

Beyond the five core agents, the Marketplace now includes HubSpot-built specialists like the Deal Loss Agent, Customer Health Agent, Customer Handoff Agent, Social Post Agent, Company Research Agent, Cross-sell/Upsell Agent (beta), RFP Agent (beta), Brand Assistant, Audit Analyzer Assistant, and Shopify Store Performance Agent (beta).

There are also partner-built connector apps that surface as tools and knowledge sources rather than full standalone agents — Confluence for Breeze, Asana for Breeze, Productboard, Guru, Helpjuice, Google Drive, Microsoft Teams, and Outlook for Breeze.

And there are partner-built Agent Tools (custom workflow actions that Breeze Agents can call). Early launch partners published by HubSpot include Sendoso, hapily, Portant, and Weave + Blend. Per HubSpot's own developer case study on Sendoso: "customers are sending gifts up to 40 percent faster and Sendoso has seen a 30 percent month-over-month increase in gift sends within Hubspot since releasing its agents."

Pricing model on Marketplace agents: mostly free to install. You pay only when they run — in HubSpot Credits, or per outcome on the two main agents (Customer Agent at $0.50 per resolved conversation, Prospecting Agent at $1 per recommended lead, both since April 14, 2026). Some partner connector apps need their own underlying subscription.

Audit Cards — the feature that makes Breeze acceptable in regulated industries

An Audit Card is a per-conversation record showing exactly which CRM properties a Breeze agent modified, what the lead-qualification outcome was, and which data points the agent used to reach that determination. They rolled out in January 2026.

HubSpot's product update copy: "an 'audit card' shows what actions the agent took in order to answer the question or when it performs actions like identifying customers for CRM and when the agent has discovered a lead… The card displays which specific properties were modified and provides transparency into the agent's CRM actions."

This is the timestamped audit trail compliance teams demand. It is live now for the Customer Agent and expanding to other agent surfaces.

The adjacent governance levers worth setting on day one:

AI Settings (Super Admin) — turn generative AI on or off, control which data classes (CRM, files, conversations, calls) Breeze can use.

HubSpot Credits cap — set a monthly spending limit. This is the single most important guardrail on an Enterprise rollout where multiple teams are building agents simultaneously.

Centralized audit log — every create/edit/publish/delete in HubSpot, including from agents, is logged.

Note. Audit Cards are what unblock Breeze in regulated industries — financial services, healthcare, public sector. If you're selling Breeze into a compliance-conscious buyer, this is the demo to lead with.

What it now costs to build AI agents in HubSpot

HubSpot Credits are the metered currency that pays for AI usage across Breeze, Smart Properties, and AI workflow actions. They reset monthly and do not roll over.

Per the HubSpot Knowledge Base "Manage HubSpot Credits," monthly allocations are:

Pricing benchmarks worth memorising:

• 10 credits — a single Data Agent prompt, a Smart Property fill, or an AI workflow action.
• 100 credits — a full Customer Agent conversation (before the April 14 outcome-pricing shift on Customer and Prospecting Agents).
• $0.50 — per resolved Customer Agent conversation (outcome-priced since April 14, 2026).
• $1.00 — per recommended Prospecting Agent lead (outcome-priced since April 14, 2026).
• ~$10 per 1,000 credits — additional credit pack pricing, or roughly $9 per 1,000 on annual.

The credit pool is shared. Heavy Smart Property runs or Prospecting Agent volume can starve Customer Agent conversations on the same account.

Pro Tip. On day one of every engagement, set a HubSpot Credits cap. Cap it at 90% of the monthly allocation by default, with overage notifications turned on. This single setting prevents 80% of the "credit blew up overnight" support tickets.

Part 2 — The Developer Track

This is the layer your engineering-capable customers and Superwork's technical consultants live in.

There is no single "Breeze Agent API endpoint" you POST to. Breeze agents are invoked through the Studio UI, the Run Agent workflow action, Breeze Assistant, or the channels the Customer Agent is wired into (chat, email, and as of January 2026, SMS, Instagram, Telegram, LINE, WhatsApp, and Slack).

Developer extensibility runs through five surfaces.

Surface 1: The remote HubSpot MCP server

The remote HubSpot MCP server at mcp.hubspot.com is HubSpot's official Model Context Protocol endpoint, exposing CRM and engagement data to any MCP-compatible AI client. It went generally available on April 13, 2026, graduating from public beta.

It is the bridge between any MCP-compatible AI client (Claude Desktop, Claude Code, ChatGPT, Cursor, VS Code, Windsurf) and the customer's HubSpot CRM.

What's exposed at GA:

• Read/write on CRM objects — contacts, companies, deals, tickets, carts, products, orders, line items, invoices, quotes, subscriptions, segments (lists).
• Read/write on engagements — calls, emails, meetings, notes, tasks.
• Read-only on marketing/content objects — campaigns and campaign metrics, landing pages, website pages, blog posts.
• Read on associations between any of the above.

Auth is OAuth 2.1 with PKCE — mandatory. The HubSpot admin sets up an MCP auth app under Development → MCP Auth Apps, generates OAuth credentials, and connects the client. Scopes are derived from the tools the MCP server exposes and the permissions of the connecting user. Existing HubSpot permissions are respected throughout.

Rate limits follow standard private-app limits. Pro: 190 req/10 s and 650K/day. Enterprise: 190 req/10 s and 1M/day.

The remote MCP server is built on the CRM Search API. It does not include vector search.

Surface 2: The Developer MCP server (local, CLI)

The HubSpot Developer MCP server is a local CLI-based MCP server that gives agentic IDEs HubSpot-specific developer context — for building apps, modules, serverless functions, and CMS content. It went GA on February 19, 2026. Different server, different job from the remote MCP server.

You install it via hs mcp setup and run it locally inside Cursor, Claude Code, OpenAI Codex, Gemini CLI, or any other MCP-compatible agentic IDE. HubSpot CLI version 8.2.0 or higher is required.

This server is for HubSpot developers building apps and CMS content. It gives the IDE HubSpot-specific context — scaffolding projects, generating modules and serverless functions, surfacing build errors, searching HubSpot dev docs in-IDE.

Note. Two MCP servers. Two jobs. The remote one connects AI to your CRM data. The local one connects AI to HubSpot's developer platform. Don't conflate them in a client conversation.

Surface 3: Customer-connectable MCP servers inside Breeze Studio

Breeze Studio includes a built-in MCP client, which lets your custom agents call out to external MCP servers as tools — Notion, Atlassian, Asana, Zapier, G2, Linear, Gong, Amplitude, and others.

Currently supported third-party MCP servers, per the HubSpot KB (last updated April 13, 2026):

• Notion — query and update databases and pages.
• Atlassian — Jira issues and Confluence content.
• Asana — tasks, projects, timelines.
• Zapier — trigger and monitor automations.
• G2 — software reviews, ratings, competitor research.
• Linear — engineering issues, projects, sprints.
• Gong — call recordings, transcripts, deal-risk and competitive-mention analysis.
• Amplitude — product analytics, user behavior, A/B tests, feature-flag performance.

Connection is OAuth (Notion, Asana) or tokenized URL (Zapier). The wire-up: clone an agent → Configure → Add tool → MCP Servers tab → Connect and add.

Bottom line. Rather than ingesting your Notion or Confluence wiki into a Knowledge Vault, give the agent an MCP tool and it queries at runtime. Cleaner pattern for any data source that changes more than weekly.

Surface 4: Custom-code workflow actions (the hybrid pattern)

A custom-code workflow action is a serverless Node.js function (running on AWS Lambda) that HubSpot calls when a workflow hits the action. It's the right tool when you need more model control, more data aggregation, or more cost control than Studio gives.

The mature pattern: workflow trigger → custom code action → call OpenAI/Anthropic with HubSpot data → return action output → use it in downstream HubSpot steps.

Supported on Operations Hub Professional and Enterprise.

This is the right move when:

• You need finer model control (specific Claude or GPT version, JSON-mode, function calling) than the built-in Use a custom LLM action gives.
• You're aggregating data across systems before prompting — pulling order history from Snowflake plus a HubSpot ticket, say.
• You want to shift cost from HubSpot Credits to your own LLM bill.

It's also where you implement vector search, RAG over private corpora, structured output validation, or anything else outside HubSpot's first-party AI options.

Pro Tip. For high-volume use cases (10,000+ monthly LLM calls), custom-code workflow actions with your own OpenAI/Anthropic key almost always cost less than HubSpot Credits. The break-even is usually around 5,000 calls/month.

Surface 5: Agent Tools and UI Extensions

Two beta developer surfaces that extend Breeze beyond the no-code stack.

Agent Tools (beta) are enhanced custom workflow actions with metadata. Breeze agents discover them in the Studio tool picker and call them autonomously. Built with HubSpot's developer projects framework — Developer Platform v2025.2 or 2026.03 required.

UI Extensions (beta, Sales/Service Hub Enterprise) surface custom AI features directly on CRM record pages, help-desk panels, and elsewhere. React frontend (@hubspot/ui-extensions) plus HubSpot-hosted serverless backend, packaged inside developer projects v2025.1+.

Realistic UI extension patterns for a B2B SaaS mid-market client:

• Next-best email card on the contact record — pulls recent activities via GraphQL, sends to GPT-5 with deal context, returns three draft email options.
• Account brief card on the company record — pulls Smart Property values plus recent ticket/deal activity, summarizes via Anthropic, displays in a styled tile with refresh-on-click.
• Renewal risk card on deal records — sentiment analysis on the last N call transcripts fetched from Gong via MCP, with a colored health indicator.
• Help-desk sidebar suggestion card — calls your internal RAG service over your engineering wiki, suggests responses alongside HubSpot's stock Customer Agent.

The constraints to flag: extensions render in sandboxed iframes (no DOM access), only HubSpot's component library is allowed, no client-side HTTP from React (use the serverless function), no cookies for session state.

The Knowledge Vault gap (still missing)

There is no public ingestion API for HubSpot Knowledge Vaults as of May 2026. Vaults are beta and can only be configured in-app.

Today's developer-facing workaround is indirect: push your external content into HubSpot Files, knowledge base, or a custom object via the standard CRM API → build a Segment over those records → add that Segment (and any uploaded files) to a Knowledge Vault in Studio.

This is awkward for live external data sources. The cleaner pattern remains the MCP client → external MCP server route inside Breeze Studio.

If HubSpot opens a public Knowledge Vault ingestion API, it'll surface in the Developer Platform changelog. Subscribe there.

If your team needs a working architecture for HubSpot AI agents that survives the next round of platform changes — book a working session with Superwork.

What this looks like on a real client

Five concrete builds for a Nordic mid-market B2B SaaS (50–500 employees, the Superwork ICP).

All five are buildable today on Pro or Enterprise.

1. Nordic ICP Qualifier (custom agent). Trigger: new contact created. Inputs: contact and associated company. Tools: browse the web, access CRM records, update CRM records. Knowledge: vault with ICP definition PDF + target account list. Output: writes "ICP Match" enum to the contact, posts a Slack alert via the Zapier MCP server if the account is in the top tier.

2. Customer Health Watchdog (cloned Customer Health Agent). Triggered weekly on the "Top 50 ARR" segment. Extra instructions in Norwegian for the CSM team. Pulls Gong call transcripts via the Gong MCP server. Outputs a styled note on the company record plus an at-risk Slack alert.

3. RFP Drafter (cloned RFP Agent). Vault: the master answer library (PDFs + KB articles). Input: uploaded RFP document. Tools: generate content, write to CRM. Review before running is forced on. The reviewer always approves before send.

4. Pre-meeting Brief Assistant (custom assistant). Lives in Breeze Assistant. Reads the deal record, summarises the last 3 calls/emails, surfaces the company's recent news via web search, outputs a one-page brief 30 minutes before each meeting on the rep's calendar.

5. Outbound Personalisation (hybrid). Custom-code workflow action calls Claude with the contact's job title, company description, and a Norwegian-language outbound playbook. Output is stored on a contact property used as the email body in the follow-up send. This avoids agent credit consumption at scale.

The pattern across all five: small surface, narrow knowledge, one trigger, audit-logged. Not "an AI for everything." A specific agent for a specific job.

Where the customer-built layer doesn't work yet

The honest section. Seven trade-offs to scope into the project before you start, not after.

Model choice inside Studio is limited. HubSpot picks the model per agent. Studio custom agents and most Marketplace agents default to GPT-5 as of January 12. The Customer, Prospecting, and Data Agents are on their own routing. If you need GPT-5 vs. Claude 4.7 vs. Gemini control, you're on Use a custom LLM or custom-code workflow actions — not Studio.

Knowledge Vaults have no public ingestion API. Live integrations with Notion, Confluence, or Google Drive go through the MCP-tool pattern, not the vault pattern. Treat vaults as a static document store, not a live data lake.

Run Agent is free in beta — for now. HubSpot's KB says it will switch to credit-billed at GA. Model that cost into your client engagements before you architect a Run Agent–heavy automation.

Bulk enrollment of existing records into agent triggers is not supported at publish. New records flow through. Backfill has to be scripted via API or done manually.

Smart Properties charge even on a null result. Ten credits per fill, whether the agent returned a value or not. A 5,000-record bulk run is 50,000 credits — a full month of Pro allocation gone in a single click.

Vector search isn't in the remote MCP server. If your agent needs semantic retrieval over HubSpot content, you mirror to Pinecone, Weaviate, or pgvector and sync from webhooks.

The credit pool is shared. Heavy Smart Property runs or Prospecting Agent volume can starve Customer Agent conversations on the same account. Set the credit cap on day one.

Studio is BETA. UI changes month to month. Production-capable, but expect to retrain admins after major releases.

None of these are dealbreakers. They are the trade-offs you scope into the project before you start, not after.

Tier-by-tier guidance for client conversations

What to actually sell, by HubSpot subscription tier.

Starter clients. Sell the Breeze Assistant plus a small number of custom assistants in Studio. Don't sell triggered agents — they're gated. Don't sell Marketplace install — also gated. Set expectations upfront.

Professional clients. Three plays that consistently land:

1. One cloned stock agent customized to the client's ICP and tone.
2. Two or three Smart Properties on company records — start narrow, target the top ICP segment.
3. One or two AI workflow actions — Summarize record on tickets, plus Use a custom LLM with the client's own OpenAI key to keep cost off HubSpot Credits.

Enterprise clients. Full Studio rollout. Multiple Knowledge Vaults segmented by team/brand/region. MCP integrations to whichever third-party systems the client uses — Gong has the highest leverage if they have it. Audit Card review built into the CSM weekly. At least one custom-code or developer-built component (UI extension or Agent Tool) to differentiate against teams running stock Breeze only.

A 90-day rollout that consistently works

The phased plan you can put on a slide tomorrow.

Days 1–14. Clean the CRM. Configure AI Settings. Set a HubSpot Credits cap. Identify two pilot use cases — one revenue-side, one service-side.

Days 15–45. Build pilot agents in Studio. Run them off triggers in a dev sandbox. Validate Audit Cards weekly. Spot-check accuracy on every run.

Days 45–75. Roll one agent into production for a single team — typically sales BDRs or Tier-1 support. Measure resolution rate or lead quality against the pre-Breeze baseline.

Days 75–90. Expand to a second team. Layer in one developer-built component (UI extension or Agent Tool). Set a monthly governance review.

Benchmarks that should change the plan:

• If agent accuracy in spot-checks drops below 85%, pause and rework the prompt or the vault before scaling.
• If credit consumption exceeds 80% of the included pool by day 20 of a month, throttle automations or shift heavy lifting to custom-code workflow actions on your own LLM bill.
• If MCP write actions cause more than 1% incorrect CRM updates in the first month, force Review before running on all Take action tools until accuracy stabilises.

Bottom line. Ship one agent into one team's production workflow before you talk about scaling. Most Breeze rollouts that stall do so because the second agent was built before the first agent was trusted.

FAQ: Building AI agents in HubSpot

What is Breeze Studio?

Breeze Studio is HubSpot's no-code workspace for managing AI inside the portal. It's where you customize HubSpot's stock Breeze Agents, clone them, build new assistants and agents from scratch, and install agents from the Breeze Marketplace. Available on Starter, Professional, and Enterprise plans; full functionality requires Pro or Enterprise. Still labelled BETA in the HubSpot Knowledge Base as of May 2026.

Do you need to write code to build AI agents in HubSpot?

No. On a Professional or Enterprise plan, you can build AI agents in HubSpot entirely through Breeze Studio without writing code. You configure inputs, instructions, tools, knowledge, and triggers through a visual builder. Code only becomes necessary when you need finer model control, custom integrations, vector search, or UI extensions on CRM record pages.

What's the difference between a Breeze Assistant and a Breeze Agent?

A Breeze Assistant is conversational and user-invoked — you summon it inside Breeze Assistant for one-off interactive tasks. A Breeze Agent follows defined steps with inputs, tools, and knowledge to complete structured tasks, and can run on a trigger (on a schedule, on record creation, on list enrollment). Assistants prioritize speed. Agents prioritize task completion.

How much does it cost to build AI agents in HubSpot?

Most of the building is free. Running them consumes HubSpot Credits. Monthly allocations are 500 credits on Starter, 3,000 on Professional, and 5,000 on Enterprise for Marketing/Sales/Service/Content/Smart CRM. Data Hub and Customer Platform Enterprise get 10,000. A Smart Property fill or AI workflow action costs about 10 credits. The Customer Agent is outcome-priced at $0.50 per resolved conversation since April 14, 2026; the Prospecting Agent at $1 per recommended lead.

What is the HubSpot MCP server?

The HubSpot MCP server is HubSpot's official Model Context Protocol endpoint at mcp.hubspot.com, exposing CRM and engagement data to any MCP-compatible AI client like Claude Desktop, ChatGPT, Cursor, or Claude Code. It went generally available on April 13, 2026, with read/write access to contacts, companies, deals, tickets, products, orders, quotes, segments, and engagements (calls, emails, meetings, notes, tasks). Auth is OAuth 2.1 with PKCE.

Can Breeze agents access external systems like Notion or Gong?

Yes. Breeze Studio includes a built-in MCP client that lets custom agents call out to external MCP servers as tools. Currently supported integrations include Notion, Atlassian (Jira, Confluence), Asana, Zapier, G2, Linear, Gong, and Amplitude. Connection is via OAuth or a tokenized URL, configured under Configure → Add tool → MCP Servers tab on any custom agent.

What is the Run Agent workflow action?

Run Agent is a HubSpot workflow action — currently in beta with a 500 executions/day cap — that triggers any Breeze agent (stock or custom) from inside a workflow. It accepts a configurable input payload and returns either a text response or structured fields you define (text, number, boolean, date, datetime, enumeration, phone). It's free during beta and will switch to credit-billed at GA.

Are HubSpot Knowledge Vaults available via API?

Not as of May 2026. Knowledge Vaults are beta and can only be configured inside Breeze Studio. There is no public ingestion API. Workaround: push content into HubSpot Files, knowledge base articles, or custom objects via the standard CRM API, then add those records to a vault as a Segment. For live external data, use the MCP client → external MCP server pattern instead of a vault.

The takeaway

In 2026, the interesting question is no longer "which Breeze Agents should we turn on."

It's "what should we build."

The customer-built layer is the platform. Stock Breeze is the demo.

If you're on Pro or Enterprise, you have the right to build:

• Custom assistants and agents in Studio, with your knowledge and your tools.
• AI workflow actions that ride inside any automation, including the Run Agent action that turns Studio agents into proper system primitives.
• Smart Properties powered by the Data Agent.
• MCP integrations to Gong, Notion, Asana, Linear, Confluence, and any other system that ships an MCP server.
• A full developer surface — MCP, Agent Tools, UI Extensions, custom-code workflow actions — when you need to push beyond no-code.

The trade-offs are real. Model choice is limited inside Studio. Vaults have no public ingestion API. The credit pool is shared. Run Agent is free until it isn't.

But none of those are reasons not to build. They are the architecture decisions you make upfront, on a real engagement, with someone who's done it before.

That is what we do at Superwork.

If you want a working architecture for AI agents on your HubSpot stack — what to build, what to skip, what the credit math actually looks like at your volume, and where the customer-built layer is now the most interesting place to build AI agents in HubSpot — get in touch with Superwork.