How many tools should a Salesforce MCP server show your agent?

A Salesforce MCP server that exposes every tool it has does not give an agent more capability. It spends the agent's context on descriptions the agent may never call. The useful measure of a tool surface is the set an agent actually calls, not the set it is shown, and the closest analogy from editorial measurement is citation versus mention: a tool listed in context is a mention, a tool the agent invokes is a citation. Count them separately and the case for a narrow, role-scoped surface stops being a matter of taste.

What the Salesforce MCP server documentation already says

The Salesforce DX MCP Server README, read 1 Oct 2026, describes more than 60 tools organised into 15 or more toolsets, and warns that enabling all of them "can overwhelm the LLM context". Its recommendation is to select toolsets rather than switch everything on. That server is free, Apache 2.0 licensed and runs locally on credentials already authorised through the Salesforce CLI, a real advantage for experimenting without a contract.

The hosted route points the other way on breadth. According to a third-party Claudeforce explainer published by Apex Hours, read 1 Oct 2026, the hosted Salesforce MCP server has been in beta since July 2026 with four tools: Discover, Describe, Dispatch and a read-only Dispatch. That account is not a Salesforce page, so treat the count as reported rather than confirmed. The Salesforce and Anthropic press release of 26 Aug 2026, read 1 Oct 2026, carries no prices.

Neither number is a verdict. What matters is how much of each surface an agent actually uses.

Mention versus citation in an agent's context

On every turn a model parses the full text of every tool definition loaded into its context: name, description, parameter schema. That text is the canonical form the agent receives, and it costs context whether or not any tool is called.

So a loaded tool is a mention: it occupies space and competes with the request and the org metadata. A called tool is a citation: the model selected it, filled its parameters and acted on the result. The gap between the two counts is the waste the README is warning about.

A worked example: the cost of mentions

The figures here are constructed for illustration, not measured on any product. Assume each tool definition averages 400 tokens once its description and schema are serialised.

• Loading roughly 60 tools (the README says more than 60): 60 x 400 = 24,000 tokens before the agent reads a single line of the request.
• A selected toolset of 8 tools: 8 x 400 = 3,200 tokens.
• A four-tool hosted surface: 4 x 400 = 1,600 tokens.

Now the citation side. Suppose an admin team logs 200 agent sessions over two weeks with all 60 tools loaded, and 11 distinct tools are ever called. That is a citation rate of 11 / 60 = 18.3 percent. The other 49 tools were mentioned in every session and cited in none. At 400 tokens each, those 49 never-called tools cost 49 x 400 = 19,600 tokens per turn, which is 19,600 / 24,000 = 81.7 percent of the tool budget.

The denominator you choose, tools loaded or tools called, decides whether the surface looks generous or wasteful.

Why a narrow Salesforce MCP server surface is a structure decision

Breadth looks like a subject question: which Salesforce areas should an agent cover? The better question is structural: who holds which tools, at which stage, against which org? A release reviewer needs different tools from an admin drafting a Flow change, and neither needs a production write tool loaded by default.

Here governance and context budget are the same problem. A role-scoped surface shrinks each person's mention count and narrows what any single credential can ask an agent to do. FlowSprite takes that structural route: any AI the customer brings drafts a Salesforce metadata change. The change lands as a GitHub pull request in the customer's private repo, is tested in a sandbox before production, and needs a human's approval before it deploys, so production is never auto-touched. Its MCP surface is role-scoped and tier-gated, and every tool call the customer's AI executes is audited. Pricing is public, starting at EUR 99 per month for Professional (10 users), with approval workflows and separation of duties on the EUR 199 Enterprise tier, and you can see how the review-first flow is laid out before starting the 14-day trial. The honest limit: a developer who wants a free, local Salesforce MCP server to script against a scratch org is better served by the open-source DX server, and a large multi-org enterprise may find more depth in Copado Agentia, Flosum or AutoRABIT, whose own pages, read 1 Oct 2026, pitch them at enterprise-scale Salesforce release work.

What the change costs in hours

Narrowing a surface is cheap. An illustrative estimate: two hours to export a fortnight of tool-call logs and count citations per tool, one hour to map the called tools to roles, and one hour to configure the selected toolsets or role scopes. Roughly four hours for a first pass, repeated quarterly. The harder part is resisting re-adding tools "just in case": each one is a permanent mention in every session for every role that receives it.

The measurement that confirms it

Run the same session set twice, two weeks apart, and date both windows. Before the change, record tools loaded, tools called and the never-called list for each Salesforce MCP server surface in use. After the change, record the same three numbers.

The control is the never-called list. If it falls toward zero (in the illustrative example above, from 49) while task completion holds steady, the narrower surface has paid for itself. Report the result as one org and one period, not a general law: a Salesforce MCP server that fits a three-person admin team in October may decay once a fourth role arrives. Recount the citations, not the mentions, at the next quarter's review.