AI Services · Service
Claude + CRM Integration (MCP)
Connect Claude to Salesforce or HubSpot so your team can ask questions of the CRM, draft from real records and log work back, all inside the permissions you already enforce.
Short answer
Claude connects to Salesforce and HubSpot through Model Context Protocol (MCP) servers. Set up properly, each user's Claude sees only the records that user can see in the CRM, starts read-only, and writes back only through actions you approve. We pick the connection path, scope the access, build the first workflows and test them against your own records.
Most firms that buy Claude seats find the same thing in the first month: people paste CRM data into a chat window, the answers are good, and nobody can say what left the system or where it went. A governed MCP connection fixes that. Claude reads from Salesforce or HubSpot directly, under the user's own login, and every call is logged.
There are now several official ways to connect, and they behave differently. Salesforce hosted MCP servers run as the signed-in user, so sharing rules and field-level security still apply. The Salesforce in Claude plugin, announced with Claudeforce in August 2026, adds prebuilt sales skills on top. HubSpot offers a Claude connector and a remote MCP server with read and write tools. A custom MCP server or the Claude API with tool use makes sense when you need actions neither vendor ships. We help you choose, then build the version your compliance team can sign off on.
Connection paths we implement
| Path | Runs as | Best for | Watch for |
|---|---|---|---|
| Salesforce hosted MCP servers | The signed-in user (OAuth through an External Client App) | Account briefs, pipeline questions, record updates by named users | Pick read-only or read-write server configurations per group; start read-only |
| Salesforce in Claude plugin (Claudeforce) | The signed-in user, through Salesforce's AIforce layer | Prebuilt sales skills such as meeting prep and deal reviews | Pilot and beta availability in 2026; confirm for your org and edition |
| HubSpot connector and remote MCP server | The connecting HubSpot user | Contact, company, deal and ticket lookups and updates | Writes can bypass some custom validation; set write tools to require approval |
| Custom MCP server | A scoped integration user or per-user OAuth, your choice | Firm-specific actions (Apex, Flows, custodial or core-system data) | You own hosting, logging and change control |
| Claude API with tool use | Your application's service identity | Back-office automation and apps built outside the CRM | The app decides what Claude may call; validate every write in code |
| RecordIQ (Vantage accelerator) | The Salesforce user, with sharing enforced in Apex | Claude answers inside Salesforce record pages, no Agentforce license | Model usage is billed by your provider |
What you get
- Data boundary worksheet: the objects and fields each use case needs, classified with your security team
- Connection path decision per use case, written down so ownership is clear
- External Client App or connector setup, scopes, and a least-privilege permission set
- Read access to up to 5 objects and write access to 1 in the fast start; more in the Jumpstart
- Three tested workflows, for example account brief, meeting prep and activity logging
- An evaluation set of real questions with expected answers, scored before launch
- Audit logging on, a short user guide and one training session
How it comes together
- Set the data boundary. List the records each task needs and agree which Claude surface may touch them before anyone connects anything.
- Choose the path. Hosted MCP for named users, the API for applications, Agentforce for agents inside Salesforce.
- Connect read-only. Stand up the connection in a sandbox, then production, with read scopes only.
- Measure. Run the evaluation set and fix prompts, tool descriptions and permissions until the scores hold.
- Add writes behind approval. One write action at a time, each approved by a person and logged with the record ID.
Is this the right call?
Good fit when
- Your team already has Claude Team or Enterprise seats and is pasting CRM data into chats
- Reps or advisors spend real time on meeting prep, account research and follow-up notes
- You want answers across Salesforce or HubSpot plus email, files and Slack in one conversation
- Compliance wants a record of what the AI could reach and what it changed
Think twice when
- The job is deterministic routing or field updates; Flow or HubSpot workflows are cheaper and easier to audit
- Customer-facing agents must live in Service Cloud channels; start with Claude in Agentforce
- Your CRM data is full of duplicates and stale owners; fix that first with AI data foundations
Mistakes we help you avoid
Connecting as a system administrator
An admin-level connection exposes every record and every action to any prompt. Use per-user OAuth or a narrowly scoped integration user.
Trusting text inside records
Emails, web-to-case submissions and chat transcripts come from outsiders and can carry instructions aimed at an AI. Treat record text as untrusted, keep writes behind approval and test with adversarial cases.
Writing AI output straight into official fields
Land summaries and classifications in draft fields until a person confirms them, so errors don't spread into reports and forecasts.
In practice
We run Vantage Point this way. Our own team uses Claude connected to our Salesforce and HubSpot through MCP for account research, pipeline reviews and content work, with the same read-first, approve-the-writes rules we set up for clients.
Questions we get
Can Claude see records a user can't see in Salesforce?
Not when it connects through Salesforce hosted MCP servers, which authenticate as the signed-in user so sharing rules, field-level security and permissions apply. A custom integration that uses one integration user sees whatever that user sees, which can be far more than the person asking. We default to per-user authentication for assistants.
Does Claude work with HubSpot?
Yes. HubSpot offers a connector for Claude and a remote MCP server that can read and update contacts, companies, deals, tickets and custom objects. We set write tools to require approval and document what each team can reach.
How long does a Claude CRM integration take?
The CRM Connector Fast Start takes about two weeks: one connection, least-privilege access, three tested workflows and audit logging. The AI + CRM Jumpstart takes four to six weeks and adds the team rollout, more workflows and approval gates.
Will Anthropic train on our CRM data?
Anthropic states that data from its commercial services is not used to train models without the customer's permission, and zero data retention is available on request for eligible API use. Consumer accounts follow different terms, so staff shouldn't use personal accounts with client data. We confirm current terms during setup.
Free · About 3 minutes
How ready is your firm to put AI to work?
Answer 10 questions and get an instant readiness score across 10 areas, with the moves to make first for your industry, CRM and goals. No email needed to see your results.