Salesforce DX local MCP server reached end of life
On 2 November 2026 Salesforce ended support for @salesforce/mcp. What the EOL notice says, what still installs, and how teams should migrate tooling.
Salesforce's local DX Model Context Protocol server hit its published end-of-life date on 2 November 2026. The package @salesforce/mcp is deprecated on npm, the issue tracker says the repositories will be archived, and Salesforce points teams at skills instead of a local MCP process. Nothing dramatic fails at midnight: the last published build still installs. What changes is support, security fixes and the assumption that a coding agent should keep launching that package forever.
What Salesforce published
The forcedotcom/mcp README opens with an important notice: the Salesforce DX local MCP server (@salesforce/mcp) reaches end of life on 2 November 2026, with a link to issue 46 for detail.
Issue 46, filed on 5 October 2026 by a Salesforce maintainer, states the consequences in plain terms:
- After 2 November 2026 the local server is no longer maintained or supported, and no further releases will ship.
- The
forcedotcom/mcpandsalesforcecli/mcprepositories will be archived (read-only). - The npm package will be marked deprecated. It stays installable, so existing and pinned setups on
0.30.15keep working. - Salesforce tooling is now delivered as skills. The public location named in the issue is forcedotcom/sf-skills.
The npm package page for @salesforce/mcp shows the deprecation banner and lists version 0.30.15 as the published build. Deprecation is a registry warning; it does not remove the tarball. That matches the issue's note that pinned installs continue to run.
The README remains a documentation hub for the old server (configuration flags, toolsets, client examples). Treat those pages as historical once the repositories are archived. They describe a product that no longer receives releases.
What "still installable" means in practice
An end-of-life MCP server that still installs is a governance problem more than a compile error. Coding agents will keep suggesting npx -y @salesforce/mcp ... because that string appears in old docs, blog posts and training data. Developers who copy a year-old .mcp.json get a process that talks to their orgs with whatever bugs and tool surface 0.30.15 had on its last ship date.
The operational shape is familiar from other retired MCP packages: the tarball remains downloadable, so "it installed" is not evidence that it is supported. Salesforce's case is still distinct because the vendor declared EOL with an explicit skills replacement, rather than leaving an abandoned reference server without guidance.
Skills change the delivery model. Instead of a long-lived local MCP process with dozens of tools registered up front, Salesforce is steering agents toward skill folders that teach how to use the CLI and related workflows. That is a different trust and update story: skill content can move without shipping a new MCP server binary, and the agent may call the Salesforce CLI rather than a dedicated MCP stdio server. Teams should read forcedotcom/sf-skills for what is actually published today, and not assume every former MCP tool has a one-to-one skill yet.
Who needs to act
Anyone whose AI coding tools still list @salesforce/mcp (or npx args that pull it) in:
- project
.mcp.json/.vscode/mcp.jsonfiles - user-level MCP configs in Claude Code, Cursor, Copilot and similar clients
- golden images, dotfiles or internal paved-road templates
- CI jobs that start the DX MCP server for agentic checks
Org authentication itself is unchanged by the EOL notice: you still authorise orgs with the Salesforce CLI. What changes is the bridge your agents use to call DX tools.
Hosted Salesforce MCP options, if your organisation uses them, are outside the scope of issue 46, which is specifically about the local @salesforce/mcp package. Check Salesforce's documentation for which hosted options remain supported.
What it means for engineering and security leaders
EOL is a remove-or-replace ticket, not a watchlist item. Once a vendor says there will be no further releases, every remaining install is a frozen attack surface. Put the package on a blocklist for new projects and schedule removal from existing ones.
Prefer the vendor's current delivery path. Salesforce named skills (forcedotcom/sf-skills) as the replacement direction. Align internal standards with that path rather than forking the last MCP build or wrapping it yourselves unless you accept owning the security maintenance.
Watch for agent reintroduction. Even after you delete the MCP entry, an agent that "helps set up Salesforce" may recreate it. Document the approved setup in the places agents read (project instructions, skills, internal runbooks) so the default answer is skills, not npx @salesforce/mcp.
Practical takeaways
- Search the estate for
@salesforce/mcpin MCP configs and scripts. Count how many developers still launch it. - Remove or replace those entries. Follow Salesforce's skills guidance in forcedotcom/sf-skills rather than pinning
0.30.15as a long-term plan. - Update paved-road templates so new clones never ship the deprecated server.
- Tell developers explicitly that the package is end of life as of 2 November 2026, still installable, and unsupported.
- Review org credentials that were only used by the local MCP server. Rotate if the server is being retired and those tokens are broader than needed.
- Track archive status of
forcedotcom/mcpandsalesforcecli/mcpso links in internal docs can be updated when read-only mode lands.
Where shared skills and a gateway fit
The SpanDock skills documentation says an administrator can turn a skill on once so every computer gets it, and that publishing your own skills is on the Premium and Enterprise plans (starter skills are in every plan; skills are in beta). That is a way to distribute an approved Salesforce workflow skill without each developer pasting MCP JSON. Separately, the MCP gateway (Premium and Enterprise) can put organisation connectors behind one governed MCP server and, when set to Enforced, report or move aside other MCP servers on the computer. Neither page mentions Salesforce's EOL or @salesforce/mcp; they help only if you choose an approved replacement and stop the old local server from returning.