salesforce-metadata

Salesforce Metadata (salesforce-metadata-mcp)

This skill drives the salesforce-metadata-mcp MCP server (228 tools). It tells you which tools to reach for, in what order, and what to check so changes land correctly the first time.

0. Check the server is connected

Look for tools named sf_* (e.g. sf_describe_object, sf_find_tool). If none exist, the server is not installed — give the user the setup in references/setup.md and stop. This skill needs server version 3.2.1 or later; if sf_create_agent has no agentType parameter, the server is older — ask the user to update (npx -y salesforce-metadata-mcp@latest).

1. Finding the right tool (lazy toolsets)

Only the core and metadata toolsets are loaded at start, to save context. The rest (flows, apex, security, lwc, agentforce, omnistudio, devops, reports…) are one call away:

1. Don't know the tool name? Call sf_find_tool with a plain-English query ("create permission set", "activate flow"). It auto-loads the group and, for narrow matches, returns the input schema inline.
2. Know the group? Call sf_load_toolset (see sf_list_toolsets for the 34 groups). Load flows before building any Flow.
3. Tool still not callable after loading? Some clients don't refresh their tool list. Do not retry or give up — call sf_tool_schema with { "tool": "<tool name>" } to get the schema, then sf_call_tool with { "tool": "<tool name>", "arguments": { ... } }. It runs the same validation and safety checks as a direct call.

Don't guess parameter names. If you're not certain of a tool's inputs (e.g. sf_retrieve_metadata takes metadataType + componentName), call sf_tool_schema first — one cheap call beats a validation error and a retry.

Never tell the user a capability is missing until sf_find_tool has come back empty.

Use the MCP tools, not a shell. Do the work through sf_* tools — don't write local source files or switch to the Salesforce CLI. A shell deploy goes to whatever org the CLI defaults to, which may not be the org the user connected this server to.