A Salesforce MCP server for every admin and developer.
Admin, Developer, Architect, DevOps, RevOps: one Context Graph indexes your org so every role, on any MCP-compatible AI agent, works from the same map. Nine Salesforce-work tools use it today. Every one after inherits the graph.
What is a Salesforce MCP server?
A Salesforce MCP server lets an AI tool work with your org through the Model Context Protocol. Clientell AI MCP is hosted, works in Claude Code, Claude Cowork, Codex, Cursor and VS Code, and adds a context graph of your org: $249 a month for 10 people. Salesforce's own DX MCP server reached end of life on 2 November 2026.
Install steps · Documentation · Pricing · DX MCP server end of life
One MCP, one Context Graph. Every role reads and plans against the same map.
Every object, field, flow, Apex class, validation rule and permission set, indexed with live structural dependencies. Clientell MCP gives your AI agent a structural map of your org, turning complex admin, dev and RevOps questions into grounded answers.
- Objects
- Fields
- Flows
- Validation rules
- Apex
- Profiles
- Permission sets
- Reports
Salesforce Admins
Draft complex flows from plain English, trace field dependencies across the entire org, and see what a change would break before you run it.5 / 6You run it.Salesforce Developers
Give your AI client (Claude Code, Cowork, Codex, Cursor or VS Code) the org context it does not have. Trace a runtime error through every flow and field that touches it, and get metadata-grounded answers instead of guesses.2 / 6You run it.RevOps Teams
Trace what is actually driving a pipeline metric, and get a dedupe plan checked against every reference before you run it.2 / 6You run it.IT and Enterprise Architects
Audit permission set matrices, map impact before a deploy, and surface orphaned automations before they become a problem.2 / 6You run it.
It is not just a search tool. It is a living map of your Salesforce org, read from inside the AI tools your team already uses. Every role builds faster, debugs deeper, and knows before it acts.
Unified intelligence across Admins, Developers, RevOps, Architects and DevOps, powered by one hosted MCP server. IT here covers the architect and DevOps work counted on the cards.
You describe a change in plain English. A Flow that keeps failing, duplicate Accounts piling up, a report that should exist but does not. The agent plans it against your real org, and names what the change would break. You run what you approve.
How it works
Objects, fields, flows, Apex, validation rules, and permissions. One unified graph connects your entire Salesforce team to accurate org intelligence.
Complete Metadata Graph: Indexes your org structure including objects, fields, flows, Apex, validation rules, and permission sets. Kept synchronized so AI assistants never guess your schema.
Live Record Context: Reads record data on demand, read-only, through Clientell AI's auth service on your own Salesforce connection and permissions. Records pass straight to your AI client and are never copied or stored.
Safety and Governance: The AI agent plans and verifies changes against org dependencies before you deploy. Every update runs under your own credentials with a complete audit trail.
A structural map of your org, not a copy of your data
The graph indexes objects, fields, flows, and permissions: never your customer record contents. This structural separation lets your AI tools query org architecture in milliseconds while your record data stays in Salesforce.
Ask about the data boundaryDescribe the task in plain English. The MCP handles the rest.
Describe what you need in plain English. The server reads the graph, resolves cross-object dependencies, and executes the workflow end to end with zero manual configuration.
Every result comes with the change plan drafted against your real metadata. You review the output and execute it through your own CLI.
Flow authoringAdmin · Developer
Draft a new Flow from plain language, or point at one that's failing and get the actual fix.
build_or_fix_salesforce_flowNames, identifiers and relationships. Not record contents.
The graph answers which: which flow touches which field, which profile reaches which object. What it does not store: customer record contents, Slack message bodies, Gmail message bodies, Jira issue contents. Those stay in their source systems.
Read-only, and the database enforces it
Not read-only by convention. Queries against your org’s graph run through the database’s read-only execution path, so a write cannot be issued even by mistake. The graph stays a faithful index of Salesforce architecture and nothing else.
MATCH (f:Field {name: "Custom_Segment__c"})
<-[:REFERENCES]-(x)
RETURN x.type, x.name
✓ 7 rows
No write path exists on this connection
It will never touch your org without you running it
A system that cannot deploy independently has no incentive to hallucinate confidence. The agent writes the change plan, and you run it as sf commands on your machine under your own credentials. Clientell AI holds a token to build the metadata index, but nothing that can read your customer records.
Proposed
Add Segment__c to the Enterprise Lead routing decision
sf project deploy start \
--source-dir force-app/main/default/flows \
--target-org production --dry-run
Salesforce ships its own MCP server. Keep it.
Salesforce’s official MCP server is the right tool for single-record operations. Clientell AI handles the org audits, pre-deployment blast radius checks, and schema reasoning no live API call can resolve. An MCP client runs both servers side by side without conflict.
Standard Actions
single-record CRUD & live SOQL
Deep Reasoning & Audit
org graph, audits, blast radius
Workspace Conflict
coexists side by side in client
| You ask | Salesforce Official MCP | Clientell AI Context Graph |
|---|---|---|
| Audit technical debt & orphaned metadata | No API exists. Requires manual Setup inspection across thousands of items. | Generates full org debt scores, orphaned automations, and unused field reports. |
| What breaks if I deploy this change? | No pre-deploy dependency check. Deploy failures happen at runtime. | Simulates deploy blast radius across Apex, Flows, and Profiles beforehand. |
| What breaks if I delete this field? | No API call returns this. Requires manual 'Where is this used?' clicks. | Traverses the complete dependency graph in milliseconds to identify all references. |
| Draft or debug a complex Flow | Can create simple records, but cannot author or trace Flow logic. | Drafts valid Flow XML and isolates failing steps with full schema context. |
| Safe duplicate merge plan | Standard merge without child record or foreign key safety validation. | Verifies all foreign key and master-detail references before drafting the merge plan. |
| Which profiles & permission sets can edit this? | Raw metadata read, but resolving effective user access requires manual calculation. | Instantly resolves effective field and object access across all permission sets. |
| Why was this custom rule created? | Out of scope. Salesforce stores current configuration, not decision history. | Traces origin rationale from linked Jira tickets, Slack threads, and Gmail discussions. |
| Create or update records directly | Yes. Direct write capability against the live org. | Drafts the plan for your review. You execute via your local sf CLI session. |
Both servers register under different namespaces and run side by side without interference.
Compared
What each MCP server actually does
Three hosted MCP servers for Salesforce, compared on what their own documentation says. Checked 18 August 2026. Where a competitor’s docs do not cover something, this says so rather than claiming they cannot do it.
| Capability | Context GraphClientell MCP | Salesforce Hosted MCPdeveloper.salesforce.com | Cirra AIcirra.ai docs |
|---|---|---|---|
| Live org-wide dependency graph | YesObjects, fields, Flows, Apex, validation rules, profiles and permission sets, plus the links between them. | No | No |
| Writes to your org | No, by designIt reads, plans and drafts. You run the change. | YesSeparately scoped SObject servers, including sobject-all for full create, read, update and delete. | YesDocumented create, update and delete on both records and metadata. |
| Impact before a change | YesTraces what references a field or Flow and names what the change would break. | No | No |
| How access runs | Your own connectionRemote Streamable HTTP context index. Live record reads run read-only on your own Salesforce connection and permissions, and only you can use it. Deploys run through your local sf CLI. The graph stores structure, never record contents. | Salesforce's infrastructureRuns inside Salesforce's hosted infrastructure, under the authenticated user's own Salesforce permissions. | Vendor-hostedA vendor-hosted remote endpoint connected to Salesforce through OAuth. |
| Tools exposed | 9Nine Salesforce-work tools, 22 in total: metadata search, best practices, flow build/fix, docs search, org reports, work context, cross-system person lookup, duplicate-merge planning, and native report and dashboard authoring. The rest are setup, billing and connector plumbing. | VariesSplit across several scoped servers, so a single number does not describe it. | Generic toolsGeneric tools for basic Salesforce CRUD. |
| Client support | Claude Code, Claude Cowork, Codex, Cursor, VS CodeWorks in all five. Other MCP clients can connect too. | Claude, ChatGPT, Cursor, PostmanOfficial docs list these as tested, using OAuth Authorization Code with PKCE. | Claude, ChatGPT, Cursor, VS Code, GeminiOfficial docs publish setup for Claude/Claude Code/Desktop, ChatGPT, Codex, VS Code, Cursor, Cline and Gemini Enterprise. |
Salesforce Hosted MCP from its overview and server reference. Cirra AI from its public documentation. Clientell AI’s tool count from its own registered tool surface. “Not documented” means the reviewed public docs did not establish it, not that the product cannot do it. Checked 18 August 2026.
For developers
Clientell MCP and the Salesforce DX MCP server
Different layers, and you can run both. DX MCP moves code between your project and an org. Clientell AI answers what that code is connected to before you move it. An MCP client can run both servers at once, which is the setup we would expect a developer to want.
Local, open source
Salesforce DX MCP
A build-and-ship toolchain, driven by the source in your project.
- Runs with npx against an org your Salesforce CLI is already authenticated to.
- More than 60 tools across the development lifecycle.
- Metadata retrieve and deploy, including Apex.
- SOQL, permission set assignment, Apex test runs.
- LWC creation and migration, DevOps Center, Code Analyzer.
Hosted, structural index
Context GraphClientell MCP
A map of the org as it actually is, read before anything moves.
- No source tree and no deploy step.
- Indexes objects, fields, Flows, Apex, rules and permissions.
- Answers what references this field, and what a change would break.
- Reads records live, read-only, on your own Salesforce connection.
- Plans and drafts. You run the change.
One thing to be clear about. No general record CRUD tool was found in the DX MCP repo, and no live dependency graph was documented there either. Both statements are about what the repo publishes, not about what the tool could be made to do.
What we can actually show you,
and where each number comes from.
- 86.5%
- mean graded score with the graph (110 questions, Opus-graded)
- 44.2%
- same model, same questions, no graph (the control)
- 86.0%
- graph + text hybrid, indexing prose added nothing
- Partner
- Salesforce consulting partner
- 0
- changes to your org you did not run yourself
- AI-led
- managed Salesforce services available alongside MCP
How the 86.5% and 44.2% were measured: read the benchmark method
Set up and fix
Powered by
The same context layer, exposed over MCP.
The graph behind this server is not Salesforce-only plumbing. It is Clientell AI’s context layer, a drop-in layer that maps data lakes, CRMs and business systems into one governed knowledge graph, so an agent can resolve what connects to what before it acts. MCP is one of four ways to attach to it: an SDK to build against directly, this MCP server, a plugin for assistants you already run, and a gateway for routing at the org level. Same graph behind all four, so the answer does not change with the door you came in through.
Read how the context layer worksWatch where each one stops
Five recordings of Clientell AI, the agent that acts on the org. Each one reads the org first, names the fields it checked, and lays out a plan. Clientell AI asks before it changes anything in production. MCP is the layer underneath that reading and that plan, and it stops at the same line every time.
- It checks firstScans the org and checks every field, picklist value and object exists before it proposes a single screen.
- You approve“The plan looks good go ahead and deploy it to my Sandbox.”
- Then it actsThe agent builds the Flow and ships it to a Sandbox.
Built for whoever owns the org
Admins, developers, architects, DevOps and RevOps all ask the same org different questions. They get paged when a flow breaks, and they are the ones who have to explain why it was there in the first place.
- SOLO ADMIN
- REVOPS LEADS
- SF ARCHITECT
- DEVOPS TEAMS
- ORG ARCHITECT
- SYSTEM ADMINS
- PARTNER TEAMS
- ENTERPRISE OPS
- SALESFORCE OPS
- SALESFORCE TEAMS
One admin holding a decade of customization together, or a partner team walking into a client org they have never seen. Same problem: no one can tell you what depends on what.
The short guide
What an MCP server for Salesforce does
MCP stands for Model Context Protocol. It is an open standard that lets an AI tool call outside tools. A Salesforce MCP server is the Salesforce side of that link. The AI asks a question, the server runs it against your org, and the answer comes back. Without one, the AI only knows what you paste into the chat.
Salesforce MCP options in 2026
There are three ways to give an AI tool access to Salesforce today. They sit at different layers, and you can run more than one in the same client. Last checked 4 Nov 2026.
Salesforce DX MCP server
What it isA local server that Salesforce ships for developers. It runs with npx against an org your Salesforce CLI is signed in to, and covers deploys, tests, metadata and queries.
Cost and statusFree and open source. Reached end of life on 2 Nov 2026: no more releases or support. The npm package stays installable.
Salesforce Hosted MCP
What it isServers that Salesforce runs for you, generally available since 29 Apr 2026. They give an AI client records, queries and actions, and every call runs as the signed-in user.
Cost and statusIncluded with Enterprise Edition and above. Salesforce says usage will use Flex Credits once metering starts. The rate is not published yet.
Clientell MCP
What it isA hosted server with a context graph of your org: objects, fields, flows, Apex, rules and permissions, and how they link. The agent knows what touches what. Records are read live under your own permissions.
Cost and status$249 a month for 10 people and 1 production org. Sandboxes of that org are free. 14-day free trial, no card.
Sources: Salesforce hosted MCP GA announcement and the DX MCP end-of-life notice. Check Salesforce's pages for current pricing and limits. Moving off the DX server? Read what the end of life means. Weighing the hosted servers? See the full comparison.
Why the structure of your org matters
An agent that only fetches records cannot tell you what a change will break. The context graph stores what references what, so it can.
The graph holds names, IDs and relationships, never record contents, and records are read live under your own permissions. In our test, the same model scored 86.5% with the graph and 44.2% without it, on 110 metadata questions run 3 times each. Read the benchmark method and data.
Which AI tools work with Clientell MCP?
Clientell MCP works in Claude Code, Claude Cowork, Codex, Cursor and VS Code. It is a standard Streamable HTTP MCP server, so other MCP clients can connect too. See the documentation for the client list. If you use Claude and want the Salesforce side explained first, read how to connect Claude to Salesforce and what Claude Code can and cannot do for Salesforce.
How do I set up a Salesforce MCP server?
- Start the 14-day free trial. No credit card.
- Connect your Salesforce org with your own login.
- Copy your access token from your workspace.
- Add the server to your AI client with the config for that client, then ask what depends on a field in your org.
Follow the install steps, or read the full documentation. Start the trial (14-day free trial, no card).
Frequently asked
Everything you need to know about Clientell MCP
Straight answers on the graph, what it stores, and what is actually shipping.
It is the Salesforce side of the Model Context Protocol, an open standard that lets an AI tool call outside tools. The AI asks a question, the server runs it against your org, and the answer comes back. Clientell MCP is one. It gives the AI a live map of your flows, fields and permissions.
Yes. Salesforce offers hosted MCP servers for Enterprise Edition orgs and above, and Clientell MCP works in Claude Code too, along with Claude Cowork, Codex, Cursor and VS Code. You can run both servers in the same session.
Yes, two. Salesforce Hosted MCP Servers have been generally available since 29 April 2026 for Enterprise Edition orgs and above, and every call runs as the signed-in user. The Salesforce DX MCP server is a free local server for developers, and it reached end of life on 2 November 2026. Clientell AI MCP is a third option. It adds a context graph of the org, so the agent knows what touches what.
Yes, it is gone as a supported product. Salesforce says the DX MCP server (@salesforce/mcp) reached end of life on 2 November 2026. It gets no more releases or support, and its repositories are archived. The npm package stays installable, so pinned setups keep working. Salesforce now ships its developer tooling as skills in forcedotcom/sf-skills. Last checked 4 November 2026.
The Salesforce DX MCP server is free and open source. It gets no releases or support after its end of life on 2 November 2026. Salesforce Hosted MCP is included with Enterprise Edition and above, and Salesforce says usage will consume Flex Credits once metering starts. The rate is not published yet. Clientell AI MCP costs $249 a month for 10 people and 1 production org, with a 14-day free trial and no card.
It depends on which one. The Salesforce DX MCP server was free, and it is now unsupported. Salesforce Hosted MCP is included with Enterprise Edition and above, and Salesforce says usage will draw Flex Credits once metering starts. The rate is not published yet. Clientell AI MCP is $249 a month for 10 people and 1 production org, plus $25 for each extra person and $200 for each extra production org.
For a team of up to 10 people with one production org, Clientell AI MCP is $249 a month, or $2,540 a year billed annually, after a 14-day free trial with no card. Sandboxes of that org are free. If you are on Enterprise Edition, you can also try Salesforce's hosted servers beside it. Check Salesforce Help for the current Flex Credit rules before you plan a budget.
Use Salesforce Hosted MCP when you need scoped record access and actions, you are on Enterprise Edition or above, and every call should run as the signed-in user. Add a third-party server when you need something it does not give you. Clientell AI MCP adds a context graph, so the agent can say what touches a field before you change it. The two run side by side in one session, and we would start with Salesforce's for record work.
Clientell MCP works in Claude Code, Claude Cowork, Codex, Cursor and VS Code. Start the free trial, connect your Salesforce org with your own login, and copy your access token. Then add the server to your AI client with the config shown for that client in the documentation.
In our benchmark it did. On 110 metadata questions, each asked 3 times, the same model scored 86.5% with the Clientell AI context graph and 44.2% without it, graded by Claude Opus. The method and the data are on the benchmark page at getclientell.com/salesforce-mcp/benchmark.
Clientell MCP costs $249 a month for 10 people and 1 production org. Each extra person is $25 a month and each extra production org is $200 a month. The 14-day free trial needs no credit card.
Salesforce's hosted MCP server can read and write records through scoped SObject servers, and a beta Headless 360 server adds some configuration actions. Cirra AI is a write-capable action layer for records and metadata. Neither publishes a live dependency graph. Clientell AI indexes objects, fields, flows, Apex and validation rules, plus the links between them, and only drafts changes for you to run. That is what lets it name a break before you run the change.
No. The graph indexes names, IDs and relationships, never record contents. Record data is read live when you ask, through Clientell AI's auth service on your own Salesforce connection and permissions. It passes straight to your AI client and is never stored.
The graph stores which component references which. When a tool drafts a change, it checks which flows, Apex or validation rules touch those fields and says so. It never applies the change itself. You run it.
Clientell MCP works in Claude Code, Claude Cowork, Codex, Cursor and VS Code. It is a standard Streamable HTTP MCP server, so other MCP-compatible clients can connect too.
Start the 14-day free trial (no credit card), connect your Salesforce org, and copy your access token. After that, adding the server to your AI client takes about a minute. The install page has the exact steps for Claude Code.
Yes. You set an active org for the session and every result names the org it came from. Developer orgs, sandboxes and production all work the same way, and each org is indexed separately.
Any MCP-compatible client can connect
Connects instantly with a single secure API token, no local server setup required. Works in Claude Code, Claude Cowork, Codex, Cursor and VS Code, and other MCP-compatible clients can connect too.
Built by the team behind Clientell AI’s Salesforce agent, for admins and teams who work in Salesforce.