What is the HubSpot Agent CLI — and what should your ops team automate first?

On May 27, HubSpot announced the Agent CLI - a new way for AI agents to operate HubSpot from the environments where ops teams actually compose their work: Claude Cowork, Claude Code, and Codex. It's in private beta, and most of the coverage so far repeats the announcement.

This article is different, because we didn't read about this way of working - we already work this way, every day. The website you're reading was rebuilt, audited, and is now content-managed by an AI agent working alongside our founder in Claude Cowork: blog posts updated through HubSpot's APIs, weekly visual QA running on a schedule, hygiene checks before every deploy. So when HubSpot says agents should be able to "run HubSpot," we can tell you concretely what that feels like - and what your team should automate first.

What exactly did HubSpot announce?

The Agent CLI is the third leg of HubSpot's agent infrastructure. The first two already exist: the public APIs (programmatic access) and the MCP server / AI Connectors (which let tools like Claude and ChatGPT read your CRM in conversation). The CLI adds the piece built for work that runs without a human in the loop: repetitive, bulk, and scheduled jobs that agents execute on their own infrastructure - before you even get to your desk, as HubSpot puts it.

HubSpot's examples are telling: a Monday 8 a.m. report of high-fit contacts with missing enrichment, a daily scan for deals closing this week with no recent activity, an automated account review per customer, a ticket-pattern flag for top-tier accounts. Notice the shape: none of these are "chat with my CRM." They're standing orders.

Why does this matter for B2B teams (and not just developers)?

Because the bottleneck in most RevOps setups isn't capability - it's cadence. Every portal has reports nobody pulls weekly, hygiene nobody does quarterly (one practitioner we follow reported automating away 35 hours of quarterly CRM cleanup), and follow-up checks that happen only when someone remembers. Agent infrastructure turns "someone should check this regularly" into "this is checked regularly." The marginal cost of cadence drops to roughly zero.

What does agent-run HubSpot look like in practice? (Our own portal)

A few standing orders from our own operation - real examples, running today:

• Weekly visual QA. Every Thursday at 19:00, an agent sweeps our website - desktop and mobile widths, English and Hebrew - and reports regressions before clients see them.
• Content operations through the API. Our blog-refresh program updates posts in place - body, featured image, tags, meta - through HubSpot's content APIs, preserving URLs and SEO history. A human approves; an agent executes.
• Deploy guards. Before any homepage republish, an agent verifies the page's draft state (we learned the hard way that stale drafts can carry surprises) and checks live asset integrity afterward - down to the byte size of the hero video.

The honest lesson from working this way: agents amplify the quality of your setup. When our data and process were clean, the agent flew. When something was undocumented, the agent inherited the confusion. Which leads to the real question:

What should your ops team automate first?

Start with jobs that are high-frequency, low-judgment, and easy to verify:

1. The Monday hygiene report. Contacts with missing critical fields, deals with no associated company, leads sitting unassigned. One list, prioritized, delivered before standup.
2. The stale-deal scan. Deals closing this month with no activity in 5+ days - the single highest-leverage daily check in most pipelines.
3. The pre-meeting brief. Account summary, open deals, recent emails, last tickets - generated for tomorrow's calendar, automatically.
4. The post-campaign pull. Same campaign report you build by hand every month, on a schedule.

Hold off, at first, on anything that writes to customer-facing surfaces without review - emails, deal stages, public content. Our rule in practice: agents draft and check; humans approve what ships. Graduate write-access one workflow at a time, the same "assist first, then agentic" pattern HubSpot recommends for Breeze.

How does this connect to Breeze?

Breeze agents live inside HubSpot - Customer Agent on your chat, Data Agent on enrichment. The Agent CLI is for agents that live outside - in Cowork, Code, or Codex - and reach in. They're complements: Breeze handles in-app, real-time work; CLI-driven agents handle the composed, scheduled, cross-system work. A mature 2026 portal will run both.

What should you do this quarter?

Three moves: (1) join the private beta; (2) write down your team's five most repeated CRM questions - those are your first standing orders; (3) get your data and permissions agent-ready, because every agent inherits them. That third one is where most teams need help - and it's exactly what our 30-minute portal audit looks at: whether your portal is structured well enough for agents (and AI in general) to make it sing.

Frequently asked questions

Is the HubSpot Agent CLI available now?

It's in private beta as of late May 2026, with public sign-up for the waitlist. It runs on the same foundation as HubSpot's public APIs and MCP server.

Do I need developers to use it?

You need someone comfortable composing agent workflows in tools like Claude Cowork or Codex - increasingly an ops skill, not an engineering one. The harder prerequisite is a clean, well-permissioned portal.

Is it safe to let agents change CRM data?

Treat agents like new team members: scoped permissions, read-only first, audited writes, human approval for anything customer-facing. HubSpot's permission model applies to agents the same way it applies to people.

Agent CLI vs. AI Connectors - which do I need?

Connectors (MCP) are for conversations: asking questions, exploring data, human in the loop. The CLI is for standing orders: scheduled, repetitive, bulk work without a human in the loop. Most teams will end up using both.