TL;DR — Key Takeaways
- What is it? Agentforce is Salesforce's platform for building governed AI agents inside the CRM; Claude is Anthropic's AI model family and assistant — and since late 2025, the two are deeply integrated rather than competing.
- Key benefit: One AI strategy with two runtimes — production agents inside Salesforce's trust boundary, cross-application reasoning in Claude — connected by a single open standard, the Model Context Protocol (MCP).
- Cost / Investment: Varies. Agentforce is consumption-based; Claude is seat-based (Team/Enterprise). Most mid-market organizations end up running both.
- Best For: Organizations on Salesforce that also run a broader application stack — email, chat, marketing, finance — and want AI that spans all of it without giving up governance.
- Bottom Line: Stop asking "Agentforce or Claude?" Start asking "where should this agent live?" The answer to that question tells you which runtime to build on.
Every enterprise software vendor now ships "agents," and buyers are left untangling what overlaps with what. The most common version of that confusion we hear: "We're on Salesforce, we have Agentforce in the roadmap, and our teams already use Claude with the Salesforce connector. Are we paying for the same thing twice?" You're not — and once you see the architecture, the two investments reinforce each other instead of colliding.
What is Agentforce?
Agentforce is Salesforce's agentic AI layer — the part of the Agentforce 360 Platform that lets you build, govern, and deploy AI agents that operate directly on your CRM. Agentforce agents reason over your Salesforce data (including Data 360, formerly Data Cloud), execute actions through Flows, Apex, and APIs, and show up in the channels where work already happens: the service console, customer portals, websites, and Slack. Every agent runs inside Salesforce's trust boundary, with the Einstein Trust Layer enforcing data masking, audit trails, and guardrails.
In short: Agentforce is where you build agents that live inside Salesforce — especially customer-facing, high-volume, repeatable workflows that need enterprise governance from day one.
What is Claude, and what does the Salesforce Claude Connector do?
Claude is Anthropic's family of frontier AI models and the assistant products built on them — claude.ai, Claude Desktop, Claude Cowork, and Claude Code. Where Agentforce embeds agents inside one platform, Claude is the surface where a knowledge worker does open-ended work across platforms.
The Salesforce Claude Connector is the bridge: Salesforce's Hosted MCP Servers, which reached general availability in April 2026, expose your org's data and actions to Claude through the Model Context Protocol. Authentication runs through an External Client App with OAuth, and — critically for governance — every call executes as the authenticated user. Profiles, permission sets, and sharing rules all hold. If a rep can't see a record in Salesforce, Claude can't see it for them either. In early 2026, Salesforce went further and shipped bi-directional MCP Apps extensions, bringing live Salesforce context (with interactive UI) directly into Claude conversations and letting outputs flow back into Salesforce.
How are Salesforce and Anthropic partnered?
This isn't a loose marketplace listing — it's one of the deepest model-provider partnerships in enterprise software, and it has compounded quickly:
- October 2025: Salesforce and Anthropic announced a major expansion of their strategic partnership. Claude became a foundational model for the Agentforce 360 Platform and the preferred model for regulated and data-sensitive industries, served via Amazon Bedrock entirely inside Salesforce's virtual private cloud. Enterprise customers like CrowdStrike and RBC Wealth Management were named as early adopters running Claude inside Salesforce's trust boundary. The companies also announced deep Claude–Slack integration and plans to bring Agentforce 360 into Claude.
- February 2026: Salesforce shipped support for Anthropic's MCP Apps — bi-directional extensions that bring Salesforce context into Claude and push Claude's outputs back into Salesforce. Salesforce was also among the first wave of partners building dedicated integrations for Claude Cowork, Anthropic's agentic desktop product for knowledge workers.
- April 2026 (TrailblazerDX): Salesforce Hosted MCP Servers went GA, Headless 360 was announced — exposing platform capabilities as MCP tools and APIs that external agents can consume without ever touching the Salesforce UI — and Agent Fabric debuted as the governance layer for multi-agent estates. Salesforce's own Agentforce Vibes IDE now ships with Claude Sonnet 4.5 as its default coding model.
- May 2026: The Data 360 MCP Server entered Developer Preview, connecting agents directly to the Salesforce data layer.
Underneath all of it sits MCP itself — the open protocol Anthropic released in late 2024 that has since become the de facto standard for connecting AI agents to tools and data, now governed under the Linux Foundation's Agentic AI Foundation with tens of millions of installs. Salesforce didn't just adopt the protocol; it rebuilt its integration strategy around it, from the developer toolchain to the AgentExchange marketplace, where partner-built MCP servers are expanding rapidly.
How do Agentforce and Claude actually work together?
The cleanest mental model is two runtimes, one protocol, shared intelligence — playing out across three architectural planes:
| Plane | What it means | Example |
|---|---|---|
| 1. Claude inside Salesforce (model layer) | Claude models power Agentforce agents via Amazon Bedrock, running entirely within Salesforce's VPC. Prompts and CRM data never leave the trust boundary; the Einstein Trust Layer governs everything. | A customer service agent in your portal reasons with Claude but operates 100% inside Salesforce governance. |
| 2. Salesforce inside Claude (tool layer) | Hosted MCP Servers and MCP Apps give Claude governed, user-scoped read/write access to your org — alongside every other system Claude connects to. | A rep asks Claude to find stalled opportunities, cross-reference their inbox, and draft follow-ups — all in one conversation. |
| 3. Agents talking to agents (orchestration layer) | Agentforce acts as an MCP client (Beta since January 2026), consuming external MCP servers through an enterprise registry with policy enforcement. Going the other direction, Agentforce actions can be triggered from within Claude, with Salesforce's execution layer providing governance. Headless 360 makes the entire platform consumable this way. | A Claude session invokes an Agentforce quoting agent as a tool; that same Agentforce agent reaches into an external data warehouse through MCP. |
The strategic implication: you are no longer choosing a walled garden. The boundary between "Salesforce AI" and "everything-else AI" is now permeable in both directions — and the same Claude models can be doing the reasoning at both ends.
When should you build in Agentforce vs. work in Claude?
Use the workflow, not the vendor, to decide. Five questions settle almost every case:
| Question | Points to Agentforce | Points to Claude |
|---|---|---|
| Who is the user? | Customers, or employees working inside the CRM | Knowledge workers spanning many systems |
| Where does the work live? | Inside Salesforce objects, channels, and processes | Across the whole stack — CRM, email, chat, docs, finance |
| What's the volume and shape? | High-volume, repeatable, deterministic (topics, actions, guardrails) | Ad hoc, open-ended, judgment-heavy |
| What's the governance boundary? | Data must never leave the Salesforce trust boundary | User-scoped OAuth access across systems is acceptable |
| What's the output? | An executed CRM workflow (case closed, lead routed, quote generated) | Analysis, synthesis, drafts, decisions, deliverables |
Service deflection, lead qualification, order status, renewal motions — those are Agentforce builds. Pipeline analysis, account research, meeting prep, QBR decks, cross-system reporting — that's Claude with connectors. And because the planes compose, the "either/or" cases keep shrinking: the Claude-side knowledge worker can hand a task to a governed Agentforce agent mid-conversation, and the Agentforce agent can pull external context through MCP without custom integration code.
What does this look like in a real organization?
Picture a 300-person B2B company running Salesforce alongside a marketing platform, Slack, a finance system, and shared documents.
Always-on, inside Salesforce: An Agentforce service agent answers order and account questions in the customer portal 24/7, grounded in CRM and Data 360 records, escalating to humans on defined triggers. A second agent qualifies inbound leads overnight and routes them with full audit trails. Both can run on Claude models — inside Salesforce's trust boundary the entire time.
On-demand, inside Claude: A sales rep starts the morning asking Claude to surface opportunities with no activity in ten days, cross-reference recent email threads, and draft tailored follow-ups — Claude reads Salesforce through the hosted MCP server, reads email through its own connector, and writes tasks back to the CRM as that rep, with that rep's permissions. Meanwhile, an ops lead has Claude assemble the monthly business review: pipeline data from Salesforce, campaign performance from the marketing platform, spend from finance — one reasoning loop across systems no single platform can see end to end.
Same organization, same intelligence, two runtimes — each doing what it's structurally best at.
How Vantage Point approaches dual-runtime AI architecture
The market trend is unmistakable: AI strategy is becoming architecture strategy, and the consulting industry is consolidating around it. Large systems integrators keep getting acquired and rolled up on roughly a four-year cycle — and every cycle, delivery teams churn, institutional knowledge walks, and clients get re-staffed with junior benches mid-engagement. That model was already strained; in an agentic landscape that changes quarterly, it breaks.
Vantage Point takes the opposite approach: senior-only consultants, AI-augmented delivery, and dual-platform depth across Salesforce and HubSpot, with active work in Anthropic's partner ecosystem. We design where each workflow should live before anyone builds — Agentforce for governed, in-platform agents; Claude for cross-application knowledge work; MCP as the connective tissue — using our VALUE methodology (Vision, Adaptability, Leverage, User-Centric, Excellence) to keep the architecture anchored to business outcomes rather than vendor roadmaps. Because we're vendor-agnostic across both runtimes, our recommendation is the one that fits your workflow, not the one that fits a reseller quota.
Common pitfalls
- Forcing every use case through one runtime. Teams that buy Agentforce and route ad hoc knowledge work through it get frustrated; teams that try to run customer-facing volume through a general assistant get burned. Match the workflow to the runtime.
- Sloppy permission models, exposed faster. Claude's Salesforce access runs as the user — which means a loose sharing model that nobody noticed in the UI becomes very noticeable when AI can query it conversationally. Audit sharing rules and field-level security before you scale connector access.
- Building customer-facing agents outside the trust boundary. If the data shouldn't leave Salesforce, build in Agentforce with Claude as the model — don't wire a generic chatbot to raw APIs and recreate governance by hand.
- Ignoring the registry layer. Shadow MCP connectors are the new shadow IT. Use Agent Fabric and Salesforce's MCP server registry to control which agents connect to which tools under which policies.
- Skipping data quality. Both runtimes amplify whatever they're grounded in. Duplicate accounts, stale stages, and empty fields don't just look bad in dashboards anymore — they produce confidently wrong agent behavior at scale.
FAQ
Is Agentforce built on Claude?
Claude is a foundational model for the Agentforce 360 Platform and the preferred model for regulated and data-sensitive industries, delivered via Amazon Bedrock inside Salesforce's virtual private cloud. Agentforce supports multiple model providers, but the Anthropic partnership is distinctive because Claude runs fully within Salesforce's trust boundary rather than calling out to an external API.
Does Claude replace Agentforce, or does Agentforce replace Claude?
Neither. Agentforce is the runtime for governed agents that live inside Salesforce; Claude is the runtime for cross-application knowledge work. They share models, they share the MCP protocol, and they can invoke each other. Most organizations that adopt one seriously end up running both.
Is my Salesforce data safe when Claude accesses it through MCP?
The Hosted MCP Servers authenticate through an OAuth External Client App and execute every request as the signed-in user, so profiles, permission sets, and sharing rules apply exactly as they do in the Salesforce UI. Separately, when Claude powers Agentforce agents, the model runs inside Salesforce's VPC and the Einstein Trust Layer governs the interaction end to end.
What is the Model Context Protocol (MCP)?
MCP is an open protocol, originally released by Anthropic in late 2024 and now governed under the Linux Foundation's Agentic AI Foundation, that standardizes how AI applications discover and call external tools and data sources. It's the reason one Salesforce MCP server can simultaneously serve Claude, developer IDEs, and Agentforce agents — stand the server up once, and any compliant client can use it under the same trust layer.
Can Claude trigger Agentforce agents directly?
Yes — this is exactly what the bi-directional MCP Apps integration and Headless 360 enable. Salesforce-native Agentforce actions can be invoked from within a Claude conversation, with governance enforced by Salesforce's execution layer, and Headless 360 exposes platform capabilities as MCP tools that external agents can consume without touching the UI.
Do we need Data 360 (formerly Data Cloud) to make any of this work?
No — you can start with core CRM data on both runtimes today. Data 360 matters as you scale: it's how Agentforce agents get grounded in unified, harmonized data beyond standard objects, and the Data 360 MCP Server (Developer Preview, May 2026) extends that same grounding to external agents. Treat it as a maturity step, not a prerequisite.
How should a mid-market organization sequence adoption?
Start where governance risk is lowest and value is fastest: Claude with the Salesforce connector for internal knowledge work (days to deploy, user-scoped permissions), in parallel with one narrow Agentforce pilot on a high-volume workflow like case deflection or lead qualification. Use the pilot to establish your registry, guardrail, and escalation patterns before expanding the agent estate.
Vantage Point is an employee-owned Salesforce, HubSpot, and AI implementation partner. 150+ clients, 400+ engagements, 4.71 / 5.0 average engagement rating. Talk to our team.
