Salesforce skills vs MCP: which do you need?

A skill is a folder with a SKILL.md file in it. The file has a name, a description and the instructions. It can also carry scripts and reference files. At startup an agent reads only the names and descriptions, and it loads the full text when your task matches. That is how one agent can hold dozens of skills without filling its context. The format is the open Agent Skills standard, so the same skill runs in any client that supports it.

Salesforce publishes its own at forcedotcom/sf-skills: Apex, Flow, SOQL, Lightning Web Components, permission sets and more. You add them with npx skills add forcedotcom/sf-skills, or through its Salesforce Development plugin for Claude Code or Codex. The README is blunt about stability: skills may be renamed, restructured or removed between releases. Pin a version if your team depends on one.

It reached end of life on 2 November 2026. In its notice, Salesforce says the repositories will be archived and the npm package marked deprecated. The package stays installable, so a setup pinned to 0.30.15 keeps running, and nothing new ships. Its “Use skills instead” section says Salesforce tooling is now delivered as skills. The notice points to skills and names no replacement server. The full list of what stops and what keeps working is on the end of life page.

A skill is instructions. It cannot see your org. It can tell the agent to run sf project retrieve start and how to read the result. Whatever it does to your org goes through the CLI login on your machine. And a retrieve can be partial or stale, so the agent is still guessing about anything you did not pull down. A skill cannot tell you that a validation rule nobody put in the repo will reject your new field.

That is the gap an MCP server fills. The Clientell AI MCP server keeps an index of your org’s names, identifiers, metadata and relationships. It does not hold record contents. One question, such as what touches Account.Custom_Segment__c, can cover Flows, Apex triggers, validation rules and permissions together. On a metadata-question benchmark the same model scored 86.5% with the graph and 44.2% without it. The questions and scoring are on the benchmark page. Clientell AI asks before it changes anything in production.

1. Install the Salesforce CLI v2 or later and sign in to a sandbox: sf org login web --alias myOrg. The skills expect an authenticated org.
2. Install the skills.

You can take one at a time instead, for example npx skills add Clientell-Ai/salesforce-skills@sf-deploy. The catalog includes sf-apex, sf-test, sf-flow, sf-lwc, sf-soql, sf-security, sf-deploy, sf-data and sf-schema.

3. Add the Clientell AI MCP server to your client. The steps are on the page for your client: Claude Code, Cursor, Codex, VS Code or Claude Cowork.
4. Put the order in your prompt. Ask the org first, then let the skill write.

The MCP call finds the validation rule. The skill writes Apex that will not trip it. The deploy then runs through your own sf CLI session after you approve. This is the shape of the request, not a recorded result. Keep the sandbox in the prompt until you trust the output.

If you write Apex and Flows all day and your repo is the source of truth, start with skills. They make the code better, and they cost nothing to try. If you are an admin or a consultant working in an org you did not build, start with the MCP server, because your first problem is knowing what is in there. If you deploy, you want both. We would not run a deploy off skills alone.

Do Salesforce skills replace MCP servers?

No. A skill is written guidance that an agent loads when a task matches. An MCP server gives the agent tools that reach a live system. Salesforce moved DX MCP users to skills, but a skill still has no view of your org unless something retrieves it.

Is the Salesforce DX MCP server dead?

Salesforce ended support on 2 November 2026 and said its repositories would be archived and the npm package deprecated. The package stays installable, and pinned setups on 0.30.15 keep working. No further releases ship.

Can I use skills in Cursor, Codex or Claude Code?

Yes. The Agent Skills format is an open standard, and both Salesforce's sf-skills repository and the Clientell AI skills repository install with npx skills add. The Clientell AI repo lists Claude Code, Cursor, Codex, Gemini CLI and VS Code Copilot as clients.

Does Clientell AI publish its own Salesforce skills?

Yes. github.com/Clientell-Ai/salesforce-skills is a public repository under the Apache-2.0 license. It covers Apex, tests, Flow, LWC, SOQL, security review, deployment, data and schema work, among others.

Should I install skills, the MCP server, or both?

If you mostly write code, start with skills. If you mostly answer what a change would touch in a live org, start with the MCP server. If you deploy, we would use both, and the order in the prompt matters more than which you installed first.

• What stops when the DX server ends: Salesforce DX MCP end of life.
• Salesforce’s own hosted option: Clientell AI vs Salesforce hosted MCP.
• A terminal workflow: Claude Code for Salesforce admins.
• Client setup: Clientell AI MCP documentation.