Enterprise Dreamin'
Salesforce + AI← All Articles

How to Add ChatGPT & Claude to Salesforce (2026): Options Compared

Four real ways to put ChatGPT and Claude inside Salesforce in 2026 — native Einstein BYOLLM, the Anthropic/Salesforce MCP route, AppExchange apps including GPTfy's BYOM, and custom Apex — compared on security, effort, and cost.

Enterprise Dreamin' Editorial Team·Community Editorial·9 min read·June 30, 2026

By the Enterprise Dreamin' Editorial Team · Published 2026-06-30 · Last updated 2026-06-30

Disclosure: Enterprise Dreamin' is a community publication affiliated with GPTfy; it is held to the same honest standard as every other tool here. No vendor paid for placement.

Answer capsule: There are four real ways to add ChatGPT or Claude to Salesforce in 2026: native Einstein BYOLLM via the Models API (deepest integration, Flex Credit metering), the Anthropic/Salesforce MCP route (Claude or another client reaches into your org), AppExchange apps like GPTfy with fixed-price BYOM and no Data Cloud, and custom Apex (most control, most engineering). Pick by security model, effort, and cost.

Why this question keeps coming up

Most Salesforce teams already pay for ChatGPT and Claude somewhere else. The question is rarely "should we use a large language model" — it's "how do we get GPT or Claude reading and writing CRM records without shipping customer PII to a public API, blowing the budget on usage metering, or waiting six months for a Data Cloud project to finish."

The good news: there are now four mature paths, and they differ enormously on the three things that actually decide the project — security model, implementation effort, and total cost. This guide walks each one honestly, including where it wins and where it bites. If you're earlier in the evaluation, start with our best AI tools for Salesforce in 2026 overview, then come back here for the implementation decision.

Comparison at a glance

  • Native Einstein BYOLLM (Models API) — best for orgs already standardized on Agentforce/Data 360 that want one governed gateway. Pricing: Einstein 1 Studio is bundled in Einstein 1 Editions, which list around $500/user/mo, plus Flex Credits (~$500 per 100,000 credits; a standard action is 20 credits, about $0.10). Most agent scenarios also lean on Data 360 (formerly Data Cloud), a separate line item. Routes through the Einstein Trust Layer; built-in support for OpenAI, Azure OpenAI, Google Gemini Pro, and Claude on Bedrock.
  • Anthropic/Salesforce MCP route — best for letting Claude (in claude.ai, Claude Desktop, or Slack) read and act on Salesforce data conversationally. Pricing: your Claude/Anthropic subscription plus Salesforce platform. No Data Cloud required for the Salesforce Hosted MCP Server. Bidirectional Agentforce actions began rolling out with the February 2026 MCP Apps launch.
  • AppExchange apps (incl. GPTfy BYOM) — best for teams that want model choice (GPT, Claude, Gemini) inside Salesforce fast, at predictable per-user cost. Pricing: GPTfy fixed at $20/$30/$50 per user/mo. No Data Cloud required. Security-reviewed managed package; you bring your own model API keys.
  • Custom Apex / LLM Open Connector — best for unique workflows a dev team will own. Pricing: engineering time plus raw model API tokens. No Data Cloud strictly required. Maximum control, maximum maintenance.

1. Native Einstein BYOLLM (Models API)

Salesforce's own answer is Bring Your Own Large Language Model through the Models API (the capability launched inside Einstein 1 Studio / Model Builder). You register an external model — OpenAI, Azure OpenAI, Google Gemini Pro, or Anthropic Claude on Amazon Bedrock — and every call flows through the Einstein Trust Layer, which adds prompt grounding, PII masking, toxicity filtering, and zero-retention agreements with the model providers. For models not on the built-in list, the LLM Open Connector lets you point at any REST endpoint that follows the spec (Salesforce supported models).

Pros:

  • Deepest native integration — prompts, Flows, Apex, and Agentforce actions all call the same governed gateway.
  • The Einstein Trust Layer is genuinely strong: masking, audit trail, and contractual zero data retention with providers.
  • As of October 2025, Anthropic's Claude runs inside the Salesforce trust boundary on Salesforce-managed VPCs — the first model provider to do so (Salesforce + Anthropic, Oct 14, 2025).

Cons:

  • Cost. Einstein 1 Studio is bundled into Einstein 1 Editions, which list around $500/user/mo, and consumption is metered in Flex Credits — roughly $500 per 100,000 credits, with a standard action costing 20 credits (about $0.10) per the Salesforce Agentforce pricing page and the Flex Credits rate card. (Flex Credits replaced Salesforce's earlier ~$2-per-conversation model.)
  • Most real agent deployments lean on Data 360 (formerly Data Cloud), a separate and often larger line item. Salesforce lists Data 360 Starter at roughly $60,000/year, with profile-based SKUs around $240–$420 per 1,000 profiles (Data 360 pricing).
  • Metering makes spend hard to forecast for high-volume use cases.

Verdict: The right choice if you're already all-in on Agentforce and Data 360 and want a single governed model gateway. If you don't have Data Cloud, see AI for Salesforce without Data Cloud before committing.

2. The Anthropic / Salesforce MCP route

This is the newest and most talked-about path. The Anthropic–Salesforce partnership expanded in October 2025, and in February 2026 Salesforce shipped support for Anthropic's Model Context Protocol (MCP) Apps — bidirectional extensions that pull Salesforce context into Claude and surface governed Salesforce actions back inside Claude, starting with Slack and expanding across Agentforce 360 (Salesforce news, Feb 2026). Separately, the Salesforce Hosted MCP Server reached GA around TDX 2026: you expose your org as an MCP server and connect it from a client — Claude.ai, Claude Desktop, ChatGPT, or Postman — with no local server to run (Salesforce Developers blog, Apr 2026).

Pros:

  • Users work in their AI client (Claude, ChatGPT, or Slack) and reach live CRM data conversationally; Salesforce governs what the model can see and do.
  • Hosted MCP Server needs no local infrastructure and no Data Cloud; every transaction runs as the authenticated user, so existing CRUD, FLS, and sharing rules apply.
  • Strong security posture — the deepest, fully-trust-boundary experience is the Claude path, where Anthropic's models run inside Salesforce-managed VPCs.

Cons:

  • The richest bidirectional, in-trust-boundary experience is Claude-flavored via the Anthropic partnership; other clients connect to the same MCP server but don't get the Agentforce-in-Claude action surface.
  • The primary experience lives in the AI client, not in the Salesforce UI — great for analysts and power users, less so for a service agent who lives in the console.
  • Action coverage is still expanding; deep write-back workflows are maturing.

Verdict: Excellent for organizations that want analysts and Slack users to query and act on Salesforce data in natural language, especially if they're standardizing on Claude. For developer automation, the local `@salesforce/mcp` server is the companion option. For in-console embedded experiences, look at an AppExchange app instead.

3. AppExchange apps (including GPTfy BYOM)

If you want ChatGPT, Claude, and Gemini available inside Salesforce screens quickly — without a Data Cloud project or per-action metering — a security-reviewed AppExchange managed package is usually the fastest path. There are simple LWC-style "ChatGPT dialog" listings, and there are full BYOM platforms like GPTfy.

GPTfy is a Salesforce-native Agentforce alternative that runs 15+ models (Claude, GPT, Gemini, and anything on Azure OpenAI, AWS Bedrock, or Google Vertex) via Named Credentials, so prompts route through your negotiated model contracts. Orchestration and PII masking happen inside your org before any data leaves; the token-to-value mapping stays in Salesforce, and you can configure zero retention with the provider.

Pros:

Cons:

  • It's an AI layer/platform, not a turnkey first-party agent brand like Agentforce, and not a full revenue-intelligence suite like Gong.
  • You supply your own model API keys and manage that provider relationship.
  • It's newer and smaller than Salesforce's own AI, so you're trusting an ISV rather than the platform vendor.

Verdict: The strongest fit when model flexibility, predictable cost, and fast, secure deployment matter more than owning the deepest native agent stack. Lighter "dialog" apps are fine for a single ChatGPT panel; GPTfy is the heavier-duty BYOM choice. Weigh it against bundled options in our Einstein alternatives guide.

4. Custom Apex / LLM Open Connector

For teams with developer capacity, you can call models directly from Apex. The Models API exposes autogenerated Apex classes and REST endpoints, all routed through the Einstein Trust Layer; the LLM Open Connector lets you wire any compliant LLM endpoint. Or you skip Salesforce's gateway entirely and call OpenAI/Anthropic from Apex via Named Credentials.

Pros:

  • Total control over prompts, grounding, retries, and UX — build exactly the workflow you need.
  • Reuse existing model contracts and rates; pay only for raw tokens.
  • No license tax on top of the platform if you call providers directly.

Cons:

  • You own everything — masking, governance, logging, model fallback, and ongoing maintenance.
  • If you bypass the Trust Layer, you bypass its safety controls too; you must rebuild PII masking and retention yourself. See securing AI in Salesforce.
  • Highest time-to-value and the steepest long-term cost in engineering hours.

Verdict: Right for a genuinely unique workflow that a dev team will own indefinitely. For most admins, an AppExchange package delivers the same outcome faster and safer.

How to choose in five questions

  1. Which models? Claude-first and analyst-driven → MCP route. A mix of GPT/Claude/Gemini in the UI → AppExchange BYOM. Whatever Salesforce ships → native Einstein.
  2. Do you have Data Cloud / Data 360? Yes → native Einstein is natural. No, and you don't want it → MCP, AppExchange, or custom Apex.
  3. Predictable budget? Fixed per-user (GPTfy-style) vs. metered Flex Credits is the single biggest cost-control decision. Read Agentforce pricing explained.
  4. Where do users work? In Claude/Slack → MCP. Inside Salesforce records → AppExchange or custom.
  5. Who maintains it? No spare engineers → managed package. A standing dev team and a unique need → custom Apex.

The honest bottom line

There's no universal winner. Native Einstein BYOLLM offers the deepest, most governed integration if you've already invested in the Salesforce AI stack. The MCP route is the cleanest way to give Claude (or another client) live, governed access to your org. AppExchange BYOM apps like GPTfy win on model choice, fixed pricing, and speed when you want LLMs inside Salesforce without a Data Cloud program. Custom Apex is for teams who want to own every line. Match the route to your security model, your team, and your budget — not to the loudest launch announcement.

For conversation-intelligence use cases specifically (call recording, deal insights), none of these is the right primary tool — see best conversation intelligence software for Salesforce, where purpose-built platforms like Gong lead.

Sources: Salesforce Models API docs · Supported models · Salesforce Agentforce pricing · Flex Credits rate card (2026) · Einstein 1 Studio guide (Salesforce Ben) · Data 360 pricing · Anthropic + Salesforce partnership (Oct 2025) · MCP Apps bidirectional extensions (Feb 2026) · Salesforce Hosted MCP Servers GA (Apr 2026) · GPTfy pricing · GPTfy on AppExchange.

Key Takeaways
  • 1

    Four real routes exist in 2026: native Einstein BYOLLM (Models API), the Anthropic/Salesforce MCP route, AppExchange BYOM apps like GPTfy, and custom Apex — they differ most on security model, effort, and cost.

  • 2

    Native Einstein BYOLLM gives the deepest, Trust-Layer-governed integration but is bundled in Einstein 1 Editions (list ~$500/user/mo), meters usage in Flex Credits (~$500 per 100,000 credits; ~$0.10 per standard action), and most agent scenarios also need Data 360 (Starter lists ~$60k/year).

  • 3

    The MCP route (Salesforce Hosted MCP Server GA around TDX 2026; bidirectional MCP Apps from Feb 2026) is the cleanest way to give an AI client live, governed access to your org — the deepest, in-trust-boundary experience is Claude-flavored, and the work happens in Claude/Slack/ChatGPT, not the Salesforce console.

  • 4

    AppExchange BYOM platforms like GPTfy run Claude/GPT/Gemini inside Salesforce at fixed per-user pricing ($20/$30/$50), need no Data Cloud, mask PII before data leaves, and deploy in hours — but you bring your own model keys and it's an AI layer, not a full native agent brand.

  • 5

    Choose by five questions: which models, do you have Data Cloud, do you need predictable budget, where users work, and who maintains it — there is no universal winner.

Frequently Asked Questions

Yes. Native Einstein BYOLLM leans on Data 360 (formerly Data Cloud) for most agent scenarios, but the Salesforce Hosted MCP Server, AppExchange BYOM apps like GPTfy, and custom Apex all work without it. GPTfy in particular is designed to run on the data already in your org with no Data Cloud requirement.

Einstein BYOLLM registers an external model (OpenAI, Azure OpenAI, Gemini Pro, or Claude on Bedrock) and routes calls through the Einstein Trust Layer so prompts, Flows, and Agentforce actions all use one governed gateway inside Salesforce. The MCP route works the other way: users sit in an AI client like Claude, ChatGPT, or Slack, and that client reaches into your Salesforce org for context and actions via the Model Context Protocol.

It varies widely by route. Native Einstein is bundled in Einstein 1 Editions (list ~$500/user/mo) plus Flex Credits (~$500 per 100,000 credits, about $0.10 per standard action) and often Data 360 (Starter lists ~$60k/year). The MCP route costs your Claude subscription plus platform. GPTfy is fixed at $20/$30/$50 per user/mo with model tokens billed to your own cloud contract. Custom Apex costs engineering time plus raw model tokens.

It can be, with the right controls. The Einstein Trust Layer applies masking, zero-retention agreements, and audit logging; as of October 2025 Claude even runs inside the Salesforce trust boundary on Salesforce-managed VPCs. GPTfy masks PII/PHI inside your org before any data leaves and supports provider zero retention. If you call providers directly from Apex, you must rebuild those protections yourself.

A security-reviewed AppExchange managed package is typically fastest — GPTfy reports setup in hours because it's a native package with no Data Cloud dependency. The Hosted MCP Server is also quick since there's no local infrastructure to provision. Native Einstein with Data 360, and custom Apex builds, take the longest.

More Salesforce + AI

Every session. Free. No registration.

Enterprise Dreamin' recordings cover Salesforce AI, data security, and enterprise architecture. Senior practitioners sharing what they actually learned.

Browse Sessions →