What Does It Take to Connect Salesforce MCP and the Claude Connector?

The request usually comes from sales. Can Claude or ChatGPT work directly with our Salesforce data? Salesforce keeps building its own AI into the platform but what we see is that most teams want to pull CRM data into the AI tool they already use and work with it there.

Salesforce MCP is how that connection happens. If you’ve connected HubSpot to an assistant, you know that’s close to one click. Salesforce asks for more. Whether you use the Salesforce Claude connector or link another assistant, the connection depends on your edition, an admin's setup work, and a clear view of cost. None of it is hard for a Salesforce admin, but all of it is worth understanding before a rollout.

What Salesforce MCP Is

MCP (Model Context Protocol) is an open standard that lets an AI assistant work with data in your CRM and other business systems. Salesforce's version is a set of hosted MCP servers that run on Salesforce's own infrastructure. Your team asks questions in plain language, and the assistant connects to a Salesforce MCP server to pull the records that answer them. Where an admin allows it, the assistant can also create and update records. The hosted servers have been generally available since April 2026, so this is a supported product rather than a beta.

The connection doesn't add a new layer of access. Every call runs as the signed-in user, with that user's field-level security, object permissions, and sharing rules applied. If a rep can't open a record in Salesforce, the assistant can't reach it on that rep's behalf either.

It also helps to separate MCP from Agentforce, since both come up in the same conversations. Agentforce is Salesforce's AI working inside Salesforce. MCP runs the other direction, carrying Salesforce data out to the assistant your team already uses. The two can easily coexist.

Salesforce also has separate MCP tooling for developers building on the platform, but that’s a different job from connecting CRM data to an assistant. For the connection itself, Salesforce supports each hosted server version for at least three years after its first release, which matters if you’re building workflows you expect to rely on.

Ways to Connect Salesforce to Claude or ChatGPT

Most companies connect in one of two ways.

Salesforce in Claude is a plugin Salesforce built for Claude. It's the most packaged option for sales teams, bundling the Salesforce connector, a Slack connector, and prebuilt sales skills. Salesforce approves organizations for access, so it starts with a request rather than an install.

A custom connector links an assistant directly to the hosted servers. For a Salesforce Claude integration without the plugin, an admin adds Salesforce as a custom connector in Claude. A Salesforce ChatGPT connector works the same way, set up as a custom app in ChatGPT’s developer mode, and most other tools that support MCP connect along the same lines. Some companies also place a third-party gateway between their assistants and Salesforce to manage access in one place.

Whichever path you choose, the assistant reaches Salesforce through the same hosted servers, the same app configuration, and the same per-user permissions. The real choice is which assistant your company already uses and whether the prebuilt skills matter to your team. It’s a bigger setup than the HubSpot MCP server and Claude connector, where official connectors for several assistants just come down to an admin's approval.

What Salesforce in Claude Includes

Salesforce in Claude launched as part of Claudeforce, the expanded Salesforce and Anthropic partnership announced in August 2026. The plugin ships with 37 Salesforce Claude skills, all built around sales work like renewal prep, QBR decks, pipeline coverage, and meeting follow-up. A setup skill runs the first time someone opens the plugin and tailors the rest to their role.

A skill is a packaged set of instructions for a recurring task. It tells the assistant which Salesforce data to pull and what to do with it. Skills don't grant any access beyond what the user already has. Skills also don’t have to come from a vendor. A company can build its own for the processes specific to how it sells, which is often where the value shows up.

Changes also go through an approval step. By default, Claude shows every proposed change and asks the user to allow it once, always allow it, or deny it. "Always allow" removes that confirmation step for the action. The important part is that you get control of how much or how little free rein the tools have to take specific actions.

Which Salesforce Editions Qualify

Edition is the first thing to check, before anyone spends time on setup. Hosted MCP servers are available on Enterprise Edition and above. Developer Edition orgs get them for testing.

Salesforce also says the hosted servers are intended for customers with Flex Credits, the prepaid credits Salesforce uses for usage-based billing. Customers without them can get them through their account executive or Salesforce Foundations. In practice, the gate is your edition plus a Flex Credits arrangement. We’ll come back to what that means for cost.

Salesforce in Claude adds its own requirements. Salesforce has to approve the organization, which also needs a current Sales Cloud Enterprise edition. On top of that, each user needs a paid Claude plan.

Mid-market companies often carry a mix of editions and add-ons, which is exactly where the answer gets murky. Confirm your specific edition, Flex Credits, and plugin eligibility with your Salesforce account team before planning a pilot.

Which Hosted MCP Servers to Turn On

Once eligibility is settled, the next decision is which servers to turn on. Salesforce hosted MCP servers come in standard versions, each with a fixed set of tools, and they all stay off until an admin enables them. Four of them handle CRM records.

1. SObject Reads reads and queries without changing anything. It's good for questions, pipeline reviews, and account research.
2. SObject Mutations creates and updates but can't delete. It's good for logging work and moving deals forward.
3. SObject Deletes only deletes. It's good for a supervised cleanup.
4. SObject All does everything, for trusted users who need full access.

Separate product servers cover Data 360 and Tableau Next, but for connecting an assistant to CRM data, these four are the ones that matter.

Because reads, mutations, and deletes live on separate servers, the admin decides how much the assistant can do. That choice sits on top of each user’s own permissions. The server decides what kinds of actions are possible, while the user’s Salesforce permissions decide which records and fields those actions can touch.

For most teams, we'd start with reads, add mutations for one narrow use case, and keep deletes off unless a supervised workflow needs them. Admins or partners can also publish custom tools built from Flows or Apex, either for business-specific tasks or to narrow exactly what an assistant can read.

There are no service accounts on these servers. Every action runs as a named user and appears under that user in Salesforce’s API logs, so there’s always a clear trail of who did what. One risk to keep an eye on is that records the assistant reads, like emails and notes, can contain text written to steer it. That’s one more reason to keep confirmations on for writes.

How Much Setup to Expect

For the standard path, setup is a job for a Salesforce admin, not a developer. It spans two platforms and three roles.

• Salesforce admin. Confirms the edition, turns on the chosen servers, and creates an External Client App, the configuration that lets an outside assistant authenticate to Salesforce. The admin also uses permission sets to decide who can connect. For the plugin, this admin requests access from Salesforce first.
• Assistant admin. In Claude, an organization Owner adds the connection using credentials from the Salesforce side. ChatGPT has its own equivalent step.
• Each user. Signs in once with their own Salesforce account.

None of this requires code. Because assistants can’t register themselves, every connection starts with an admin deliberately creating an External Client App, and each app can be limited to specific profiles and permission sets. We also recommend one app per assistant, so each connection can be reviewed or shut off on its own, and a sandbox test before anything touches production.

In our experience, a pilot is a short admin project, measured in hours or days rather than weeks. The longer work is deciding permissions, approval standards, and cost guardrails. Custom servers and custom tools are where developer or partner help comes in if you need it.

What Salesforce MCP Costs

Salesforce MCP pricing is better described as a structure than a price. Turning the hosted servers on is included with qualifying editions, with no separate fee. Using them is where Flex Credits come in, since Salesforce says usage may be billed.

That billing is still taking shape and changing over time. Salesforce has said that calls into production orgs from registered AI agents, meaning agents given their own identity, will be metered against Flex Credits. Customers will get advance notice before charging starts. Sandboxes and Developer Edition orgs aren't metered, so testing costs nothing. Salesforce hasn’t spelled out how metering applies to assistant connections tied to individual users, so you should plan for usage costs rather than assuming the connection is free.

Salesforce costs tend to build up across licenses and add-ons with agent usage becoming one more line to watch too. In one manufacturing engagement, Ascendex cut about $20K a year from a client's Salesforce licensing and avoided a $6.5K annual increase by assessing what they actually needed and what they were actively using. Knowing what you already pay for and use makes the new line items easier to manage.

What to Decide Before You Connect

Settling these early is far easier than changing them after people have started connecting.

1. Eligibility and cost. Confirm your edition, any add-ons, and how agent usage will be billed under your contract.
2. Path. Choose Salesforce in Claude, a custom connector, or both. Your company's AI standard usually decides this more than any feature does.
3. Servers. Decide which record servers to enable, with reads as the default.
4. Approval standard. Make "allow once" confirmations for writes a team standard, so nobody switches them off alone.
5. Pilot group. Start with a small group without admin rights, after a sandbox test.
6. Owner. Name someone to review MCP activity, which Salesforce's API logs can isolate by user. That person should also track Salesforce's frequent changes, watch Flex Credits use, and know how to reverse a bad change.

Start With Read Access

Connecting Salesforce MCP is a short project but the decisions around it determine whether the connection turns out to be useful. Start with read access for one team and a handful of real tasks. Slowly widen access once permissions, approvals, and costs are understood in real world scenarios.

Ascendex helps companies bring AI into the systems they already run, from choosing a path through rollout and the changes that keep coming after it. If you're working out how to connect Salesforce to Claude or another assistant, our CRM consulting team can help you pick the right approach and set it up with the edition, permissions, and cost controls your team needs. Reach out and connect with us if you want to have a discussion about how we can help.