Real Estate Stack · Chapter 08 of 10

Claude, set up for an agent

Everything before this chapter was choosing tools. This chapter is the one where you decide how Claude sits inside your day: which plan, which project, which connector, and which of your twelve recurring jobs it drafts first.

You don't need every Claude surface on day one. A solo agent on Pro with one Project and two connectors covers most of this chapter. Claude in Chrome, Claude Code, and Cowork are for the moment your workload actually asks for them, not before. Set the plan and the Project up once, and the twelve workflows below become copy-and-adjust, not a new prompt every time.

Which plan for which agent

Prices below are what Anthropic's own pricing page listed on 2026-09-03. verified Anything about how your chats train future models, or how long they're kept, is the same source. verified

Plan	Price (dated 2026-09-03)	What an agent gets	Data or training note
Free	$0/month	Web, iOS, Android, and desktop chat; code generation; content creation; web search; memory on by default; file creation and code execution; artifacts; skills (pre-built). Capped at 5 projects and 1 custom connector [1]	You choose whether chats train future models. Opt in and Anthropic keeps that data 5 years; opt out and it's the prior 30-day window [2]
Pro	$17/month billed annually ($200 up front), $20/month billed monthly [1]	Everything in Free, plus Claude Code, Claude Cowork, Claude Design, Claude Science, Projects, Research, more usage, and Microsoft 365 integration [1]	Same opt-in/opt-out choice as Free [2]
Max (5x / 20x)	From $100/month for 5x the usage of Pro; a higher tier for 20x [1]	Same feature set as Pro, with higher output limits, early feature access, and priority access when traffic is high [1]	Same opt-in/opt-out choice as Free and Pro [2]
Team	Standard seat $20/seat/month annual ($25/month monthly); Premium seat $100/seat/month annual ($125/month monthly, adds a Claude Code developer environment and 5x usage) [3]	All Claude features, more usage than Pro, central billing and admin, requires a minimum of two members and is not built for one person working alone [3]	Memory is off by default until an owner turns it on, and Team sits outside the August 2025 training/retention change entirely (the Claude for Work carve-out) [2] [4]

Note the two Team quirks if you're thinking about upgrading for a transaction coordinator or a second agent: it needs at least two seats, and your chats are never used for training regardless of what you pick, the opposite default from Free, Pro, and Max. verified

The office file becomes a Project

Your office file from Chapter 01 is the one document Claude should never need reminding of. A claude.ai Project turns it into permanent background knowledge instead of something pasted in every conversation. Projects are self-contained workspaces with their own chat history and knowledge base, available on every plan, though Free is capped at 5 projects. verified

In claude.ai, open Projects and choose Create project. Name it something you'll recognize in six months, for example "[Your name] Real Estate Ops."
Open Project knowledge, add the office file from Chapter 01 as a file (upload it, or connect Drive and select it, see the connectors table below).
Add a plain-text or spreadsheet list of the disclosures your brokerage and state already require you to give (agency disclosure, wire fraud advisory, and anything specific to your state from Chapter 09) so Claude never has to guess whether a form already exists for something.
Add the fair housing word list from Chapter 04 and Chapter 06 (the commonly flagged words and the safer phrasing) as its own file, so listing copy gets checked against it automatically rather than from memory.
Add your CRM's pipeline stage list from Chapter 02 (new lead, contacted, under contract, closed, whatever your stages are actually named) so Claude asks instead of assumes when a contact's stage matters.
Open Project settings and paste in project instructions. Use the block below as a start and edit the bracketed parts.

You are helping [Agent Name], a real estate agent with [Brokerage Name], licensed in [State].

Ground every answer in the attached office file, disclosure list, fair housing word list, and pipeline stage list before you improvise anything.

Never send an email or text on my behalf. Draft it, then stop and let me send it.
Never sign, initial, or submit any form for me.
Never write to the MLS or the CRM without me reviewing the change first.
Never invent a comp, a price, a date, or a fact you can't see in an attached file or a page I've shown you. Say "not found" instead.
Check every listing description or ad draft against the attached fair housing word list before you show it to me.
If a task touches housing eligibility or financial eligibility (a buyer's qualification, a CMA opinion of value, a net sheet), flag that a qualified professional needs to review it before it reaches a client, and that AI was used in drafting it.
If you're not sure which pipeline stage a contact is in, ask me. Don't guess from the list.

That instruction block is written to match the Usage Policy's own high-risk rule: legal, financial, and housing-eligibility outputs need a qualified human review before they're finalized or sent, and the use of AI has to be disclosed to the person affected. verified A CMA opinion of value or a buyer qualification note both sit inside that category, so it's worth building the reminder into the Project itself rather than trusting yourself to remember it every time.

Memory settings to check on day one

Memory behaves differently depending on plan, and the defaults are easy to miss. Check these before you start pasting client details into chats:

Memory is on by default for Free, Pro, and Max, across web, desktop, and mobile. It's off by default for Team and Enterprise until an account owner turns it on. verified
Chat search (finding an old conversation by keyword) is a paid-plan feature: Pro, Max, Team, Enterprise. verified
Memory on anything Claude flags as a sensitive topic is opt-in, and you get a notice every time it's saved. verified
Memory is shared with Cowork only when Cowork is running in the cloud, not in a local session. verified
If you're solo on Pro, decide on day one whether Claude should carry client names and details across unrelated chats, or whether everything stays inside the Project's own knowledge base with general memory off. Either is defensible; just choose on purpose.
Check the training toggle in the same settings pass. Opting in extends retention of your chats to 5 years and lets Anthropic train future models on them; opting out keeps the prior 30-day retention window. This choice doesn't apply once you're on Team or Enterprise, where it's already off by default and outside the training pipeline. verified

Connectors that matter for an agent

A connector lets Claude read and, with your approval, write to a service you already use. Free accounts get one custom connector; paid plans get more, though Anthropic hasn't published an exact per-plan cap. verified The Google connectors (Gmail, Calendar, Drive) are available on every plan, not just paid ones. verified

Connector	Claude can read	Claude can write	Approval prompt behavior	Plan needed
Gmail	Threads and messages you give it access to	Draft and send email	Approval required by default before it sends anything verified	All plans (Free through Enterprise) [5]
Calendar	Events and availability	Create or edit events	Approval required by default before it writes verified	All plans [5]
Drive	Files and folders you connect	Create or edit files	Approval required by default before it writes verified	All plans [5]
Notion	Pages and databases you connect	Create or edit pages and database rows	Listed per connector in the connectors directory, by category and read/write capability claim	Varies by workspace plan [6]
Zapier MCP	Whatever the underlying Zapier action exposes, across 9,000+ connected apps	Runs any of 30,000+ actions across those apps, including CRM and calendar writes	Standard MCP tool-approval flow; Zapier's own account controls which actions are live	Any plan that supports MCP connectors; Zapier's own pricing applies separately [7]
Make MCP	Scenario data and account contents in your Make account	Runs scenarios, manages account contents	Standard MCP tool-approval flow	Claude Desktop (MCP token) or Claude Code (OAuth); Make's own pricing applies separately [8]
HubSpot MCP	CRM contacts, deals, and records	Read and write CRM data (official vendor server, GA April 13, 2026)	Standard MCP tool-approval flow, OAuth-connected	HubSpot's own plan tiers; the MCP server itself is free on the Free CRM [9]
GoHighLevel MCP	Contacts, pipelines, calendars, conversations, tags (about 39 tools)	Write access across those same objects	Standard MCP tool-approval flow, documented for Claude in GHL's own help center	Included free on Unlimited and SaaS Pro plans [10]
Twenty MCP	CRM records in your workspace	Read and write CRM data in natural language (native server, not a third-party wrapper)	Standard MCP tool-approval flow, OAuth-connected	Ships in every Cloud workspace on paid plans [11]

If your CRM isn't on that list, check for an official MCP server before trusting a third-party one built outside the vendor. Chapter 02 has the CRM-by-CRM breakdown of who has a vendor-run server and who only has community ones.

Claude in Chrome for SkySlope and the MLS

Claude in Chrome reads the page you're signed into, then clicks, types, and fills forms while you decide what happens next. It's available on all paid plans (Pro, Max, Team, Enterprise) and is now generally available, not a preview feature. verified It's the surface closest to Claude actually operating inside SkySlope or your MLS, and the one with the sharpest guardrails.

Set up a separate Chrome profile used only for MLS and transaction-management work, with no personal email, banking, or social accounts signed in. Claude can see anything visible on that screen, so keep the screen boring.
Sign into SkySlope and your MLS (Flexmls, Matrix, whichever you use) yourself, in that profile. Never type your credentials for Claude, and never ask it to log in on your behalf.
In the extension's settings, use "ask before every action" for anything that could send, sign, or submit. Anthropic explicitly supports an "automatically approve" mode as an alternative, but that's the wrong setting for a signing envelope. verified
What it may do: fill in fields on a form it already recognizes, read an envelope's status page, rename a file using a naming convention you give it, or pull a hotsheet or saved search you already run.
What it may not do: click sign or send, enter payment details or login credentials, or try to get around a CAPTCHA. Anthropic blocks CAPTCHA bypass, stock trading and investment transactions, entering sensitive data, and scraping facial images at the extension level, and it asks permission before acting on financial sites regardless of your approval setting. verified

Anthropic's own guidance is blunt about this: running it on a signed-in account without a separate profile is "strongly discouraged," because Claude can see whatever personal or sensitive information is on screen. verified A working Arizona buyer's agent's procedure already follows this pattern: SkySlope Forms and Flexmls run in a signed-in Chrome session Claude drives one field at a time, credentials are never handled by the assistant, and a separate pane is kept for public sites like the county assessor so the MLS session stays untouched. operator

Claude Code and Cowork, for the technical agent

Most of this chapter doesn't need either of these; they matter once someone on your team wants Claude doing scheduled or scripted work rather than one drafted answer at a time.

Claude Code is a command-line and IDE agent requiring a direct Anthropic plan (Pro, Max, Team, or Enterprise). verified It integrates with the Claude in Chrome extension to test web apps, debug with console logs, automate form filling, and extract data, sharing the browser's existing login state and pausing at any CAPTCHA or login screen rather than trying to get past it. verified Its documentation is written for developers, with no published non-developer on-ramp. inference

Cowork is built for the opposite audience: it completes tasks across files and tools with oversight, opens a browser for web steps, and runs unattended or scheduled work with step-by-step visibility. verified It's in the desktop app on Pro, Max, and Team, with web and mobile rolling out in beta. verified Cowork plugins bundle skills, connectors, and sub-agents together, and Anthropic open-sourced 11 official plugins for any paid plan. verified Agent Skills, the underlying building block, run pre-built for document creation on any plan, and accept custom uploaded skills on Pro, Max, Team, and Enterprise once code execution is enabled. verified

The twelve workflows

These are the recurring jobs this guide asks Claude to do across a real estate business: trigger, tools, exact steps, and what never gets automated no matter how well it's working. The chapter number in parentheses is where the full procedure lives.

Office file to Project (Chapter 01). Trigger: you're setting Claude up for the first time, or bringing on a second agent. Tools: Projects, the Drive connector. Steps: create the Project, upload the office file, add the disclosure list, add the fair housing word list, add the pipeline stage list, paste in the project instructions block above. What stays human: the office file's contents and every fact in it come from you, not Claude.
First-five-minutes reply draft (Chapter 05). Trigger: a new lead lands in your CRM or inbox. Tools: Gmail or CRM connector, the Project. Steps: Claude reads the lead's source and form fields, drafts a personalized first reply referencing the property or search, flags anything unverified (price, availability), and stops. What stays human: you confirm the facts and send it yourself.
Weekly relationship review (Chapter 02). Trigger: a set day each week. Tools: CRM connector or export, the Project. Steps: Claude pulls contacts with no activity in 30, 60, or 90 days, groups them by pipeline stage, and drafts a short check-in note per contact. What stays human: you decide who gets contacted this week and edit each note before it goes out.
CRM import cleanup (Chapter 02). Trigger: migrating contacts into a new CRM or merging a purchased list. Tools: CRM connector or a CSV upload to the Project. Steps: Claude flags duplicates, missing consent fields, malformed phone numbers or emails, and contacts with no documented consent to text or call. What stays human: you approve every merge and deletion; Claude never writes the cleaned list back unattended.
Buyer packet in SkySlope (Chapter 03). Trigger: you tell Claude you have a new buyer. Tools: Claude in Chrome, signed into SkySlope Forms. Steps: Claude asks the buyer's name, email, phone, compensation, term, and area; creates the file; adds the standard document set; fills the known fields. What stays human: you review every field before anything sends, and you send the DigiSign envelope yourself. A working Arizona buyer's agent's procedure follows this exact pattern. operator
Under-contract date tracker (Chapter 03). Trigger: a file goes under contract. Tools: the pipeline file (SkySlope Forms holds no transaction dates), a scheduled check. Steps: Claude reads envelope status and the pipeline file, flags any buyer gone quiet or any envelope unsigned past a deadline, and reports only what needs action. What stays human: you enter contract and closing dates yourself. operator
County-record-first CMA narrative (Chapter 04). Trigger: a listing appointment or a comp request. Tools: a public-records browser pane, Claude in Chrome on the MLS, the Project. Steps: Claude verifies the subject against the county assessor first (square footage, owner of record, sale history), pulls MLS comps, then drafts the narrative reconciling the two. What stays human: you set the final price opinion; Claude's draft is a starting narrative, not the number.
Listing description with fair housing check (Chapter 06). Trigger: a new listing goes live. Tools: the Project (with the fair housing word list attached). Steps: Claude drafts the description from the property facts you give it, checks the draft against the attached word list, and flags any phrase that describes a preferred occupant rather than the property. What stays human: you make the final wording call and post it.
Showing feedback summary (Chapter 06). Trigger: a batch of showing feedback comes in. Tools: email or CRM connector, the Project. Steps: Claude reads the raw feedback, groups it into themes (price, condition, layout), and drafts a short summary for the seller. What stays human: you decide what to actually tell the seller and when.
Open house follow-up drafts (Chapter 06). Trigger: an open house sign-in sheet or app export. Tools: CRM connector, the Project. Steps: Claude drafts a personalized follow-up for each visitor based on what they said or asked about, tags likely next steps (buyer, seller, just looking), and stops. What stays human: you review and send each message; none go out unattended.
Builder comparison sheet (Chapter 07). Trigger: a buyer is choosing between builder communities. Tools: the Project, the quick-move-in data you've collected. Steps: Claude builds a side-by-side sheet of price, lot, incentives, and estimated close date across the communities considered. What stays human: you verify every number against the builder's current sheet before showing the buyer, since inventory and pricing change daily.
Monthly market update draft (Chapter 04). Trigger: the first of the month. Tools: MLS access (RPR or your own saved searches), the Project. Steps: Claude drafts a short market summary (inventory, days on market, median price movement) for your farm area or sphere, sourced from the numbers you paste in or pull. What stays human: you confirm every figure against the MLS before it goes to a client list.

What stays human

Every send: email, text, DigiSign envelope. Claude drafts; you send.
Every signature or initial on a client document.
Which Claude plan and connectors are turned on, and whether memory and training are on or off.
Any write to the MLS or CRM, even a small one.
The final price opinion on a CMA and the final wording on any listing description.
Signing into SkySlope and the MLS yourself, in a browser profile Claude in Chrome uses and nothing sensitive shares.

Do this today

Pick a plan from the table above and confirm your memory and training settings match what you decided.
Create the Project, attach the office file, disclosure list, fair housing word list, and pipeline stage list, and paste in the project instructions block.
If you'll use Claude in Chrome for SkySlope or the MLS, set up the separate browser profile first, before you connect anything else.

Sources

claude.com/pricing, retrieved 2026-09-03
anthropic.com/news/updates-to-our-consumer-terms, retrieved 2026-09-03
claude.com/pricing; support.claude.com/en/articles/9267247, retrieved 2026-09-03
support.claude.com/en/articles/11817273 (memory), retrieved 2026-09-03
support.claude.com/en/articles/11176164; support.claude.com/en/articles/10166901, retrieved 2026-09-03
support.claude.com/en/articles/11176164 (connectors directory), retrieved 2026-09-03
zapier.com/mcp; docs.zapier.com/mcp/clients/claude, retrieved 2026-09-03
developers.make.com/mcp-server; help.make.com/make-mcp-server, retrieved 2026-09-03
developers.hubspot.com/ai-tools/mcp; developers.hubspot.com/changelog/mcp-server-beta, retrieved 2026-09-03
help.gohighlevel.com/support/solutions/articles/155000008360-highlevel-mcp-multi-account-support-for-claude, retrieved 2026-09-03
twenty.com/pricing, retrieved 2026-09-03
claude.com/claude-in-chrome; support.claude.com/en/articles/12902428, retrieved 2026-09-03
code.claude.com/docs/en/chrome; code.claude.com/docs/en/overview, retrieved 2026-09-03
claude.com/cowork; claude.com/plugins; support.claude.com/en/articles/13837440, retrieved 2026-09-03
docs.claude.com/en/docs/agents-and-tools/agent-skills/overview; claude.com/blog/skills, retrieved 2026-09-03
support.claude.com/en/articles/9517075 (Projects), retrieved 2026-09-03
anthropic.com/aup (Usage Policy, high-risk use and impersonation), retrieved 2026-09-03
Operator evidence: a working Arizona buyer's agent's SkySlope buyer-packet and buyer-pipeline procedures, retrieved 2026-09-03

← Previous · All chapters · Next →

Starting from zero? Use the field manuals first.

02 · BEFORE YOU BUY SOFTWARE
If you're still choosing the buyer, offer, stack, agents, content, customers, and delivery, don't start with a tool: start with the manuals. These create local worksheets only: no checkout, outreach, or account setup.

Find the first manual

Not sure which guide to open first? Start from your stuck point and let the finder recommend the next manual and the first worksheet question. No personal data is collected: it just points you at the right starting move.

Open the guide finder →

Work the starter pack

Move through buyer, offer, stack, agents, content, customers, and delivery as a sequence of short local worksheet sessions. One artifact at a time, in order, so you build the business instead of browsing for it.

See the Business Path →

See what "done" looks like

Read a fully worked field manual before writing your own version, so the output feels concrete: a real, numbered workflow with the risky steps gated, not a fake proof claim.

Open a sample field manual →