AI Services · Service
MCP and Multi-Agent Orchestration
Connect Agentforce, Claude and OpenAI agents to each other and to Salesforce, HubSpot and outside systems, with permissions that follow your CRM and a log of every call.
Short answer
MCP, the Model Context Protocol, is an open standard that lets AI agents call tools and read data in other systems in a consistent way. Vantage Point uses it to connect Claude, ChatGPT and Agentforce to Salesforce and HubSpot, and to let one agent hand work to another, with each call running under the signed-in user's permissions, logged, and limited in what it can write.
What MCP is, in plain language
MCP gives AI agents a standard way to plug into other systems. A system publishes an MCP server that lists the tools it offers, such as search accounts or update an opportunity. An agent connects, sees those tools, and calls them as the user who signed in. Salesforce and HubSpot both publish MCP servers, so Claude, ChatGPT and other MCP clients can work with CRM data without a custom integration for each one.
Agent-to-agent orchestration
Some work crosses agents. A Claude assistant prepares a client brief from email, files and Salesforce, then hands a service request to an Agentforce agent that owns the case process. An Agentforce agent finishes intake and passes a research task to a Claude agent with longer documents to read. We design which agent owns each step, what it passes on, and what each is allowed to write, so a handoff never widens what any one agent can do.
How the pieces fit
| Connection | Runs as | Typical use |
|---|---|---|
| Salesforce hosted MCP server | The signed-in Salesforce user, through an External Client App | Claude or ChatGPT reading and updating Salesforce under existing sharing rules and field-level security |
| HubSpot MCP server | The connecting HubSpot user | Claude reading and updating HubSpot records |
| Agentforce with MCP tools | The agent's permission set | Agentforce agents calling outside systems and other agents |
| Custom MCP server | A scoped service identity | Systems without an MCP server of their own, such as a portfolio or policy system |
What you get
- A connection map: which agents, which systems, which path and why
- External Client Apps, permission sets and OAuth set up per path
- Read-only first, then write actions added one at a time behind approval
- Handoff design between Agentforce, Claude and OpenAI agents
- Logging of prompts, tool calls, results and approvers
- A test set built from real records, re-run after every change
How it comes together
- Map the work. Which jobs each agent does and which systems it needs.
- Set the boundary. Objects, fields and actions per agent, agreed with your security team.
- Connect read-only. Hosted MCP servers first, per-user sign-in, field-level security respected.
- Add writes behind approval. One action at a time, each logged with the approver.
- Orchestrate. Handoffs between agents, tested end to end.
Is this the right call?
Good fit when
- Your team already uses Claude or ChatGPT and wants CRM data in it
- You run Agentforce and need it to work with other agents or systems
- You want one governance model across several AI tools
Think twice when
- You need one agent inside Salesforce channels only; start with Agentforce
- Your CRM data isn't reliable yet; start with data readiness
Mistakes we help you avoid
Connecting with an admin account
A shared admin connection lets every user see everything. Connect per user so CRM sharing and field-level security still apply.
Turning on writes on day one
Start read-only, then add one write action at a time with a person approving each change.
Letting a handoff widen access
When one agent passes work to another, the receiving agent should never gain access the user didn't have.
In practice
NewSpring Capital. We connected Claude to NewSpring's Salesforce through Salesforce's hosted MCP server, read-only at org level, tested with a small group before wider rollout.
Health technology and manufacturing. We set up the same connection for a health technology software company, with per-user sign-in that their IT team then locked down with permission sets, and resolved a manufacturer's OAuth scope issue to get their read-only connector live.
Reference demo. An Agentforce agent handing off to a Claude agent through MCP, built by Vantage Point as a labeled demo.
Questions we get
What is MCP?
MCP, the Model Context Protocol, is an open standard created by Anthropic that lets AI agents discover and call tools in other systems. Salesforce and HubSpot both publish MCP servers.
Does MCP respect Salesforce field-level security?
With Salesforce's hosted MCP servers and per-user sign-in, the agent runs as the user, so sharing rules and field-level security apply. We avoid shared admin connections for that reason.
Can Agentforce work with Claude or OpenAI agents?
Yes. Agentforce can use Claude or OpenAI models for reasoning and can call outside tools and agents. We design the handoffs and limit what each agent can write.
Where do we start?
With an MCP Quick-Start: one connection, read-only, on your own data, with the permission model written down. Writes and agent handoffs come next.
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.