In a headless design nobody logs in. Agents call tools, and the tool call is what you pay for. Here is a burn model for MCP tool calls and completions, plus the levers to cap the cost.
MCP tool calls and model completions are the metered unit behind every Salesforce Headless 360 agent. The seat is gone. The tool call is the thing that costs. This guide gives a burn model for tool calls and completions, and the levers to cap the cost before you build.
In a headless design nobody logs in. Agents call tools. So the cost question is no longer how many seats. It is how many tool calls, at what credit draw, plus how many completion tokens.
An MCP tool call is a single invocation of one of the more than 60 tools Headless 360 exposes, and it is the point where metering attaches. The Salesforce developer blog documents the tool surface.
Each tool call is a discrete piece of work. A call that triggers an Agentforce action draws 20 Flex Credits per the Agentforce pricing page, near 10 cents.
Many tool calls read or write through Data 360, so a call carries a data processing charge as well as the credit draw. One call, two costs.
Build the burn model from three inputs multiplied at peak volume: tool calls per day, credits per call, and completion tokens per call. That gives a defensible daily cost before you commit.
Take 5,000 user requests a day, each fanning into 8 tool calls. That is 40,000 tool calls, or 800,000 credits at 20 each, near 4,000 dollars a day before the completion pass through.
Illustrative MCP tool call burn
| Tool calls per day | Credits per call | Credits per day | Approx dollars per day |
|---|---|---|---|
| 10,000 | 20 | 200,000 | 1,000 |
| 40,000 | 20 | 800,000 | 4,000 |
| 100,000 | 20 | 2,000,000 | 10,000 |
Model completions add a token cost on top of the action, because every reasoning step and generated response consumes tokens. The completion pass through stacks on the Flex Credit draw.
When an agent calls Claude Sonnet or GPT 5, the prompt and response tokens are billed through. Larger context and longer responses cost more, so prompt design is a cost lever, not just a quality lever.
A single tool call can draw a Flex Credit for the action, a Data 360 charge for the lookup, and a completion pass through for the tokens. Three costs from one call is common.
The common advice is to price agent cost per conversation, the way older chat pricing worked, because a conversation is easy to count. We disagree. In the builds we reviewed, one conversation fanned into six to twelve tool calls, and each call carried a credit, a data charge, and often a completion token cost, so per conversation math understated spend by a wide margin. The buyer side move is to price the tool call, not the conversation, model the fan out at peak, and pin the tool call unit and completion pass through as capped rates. Counting conversations is exactly how the invoice outruns the forecast.
Source: Redress Compliance advisory engagement file, 2025 to 2026.
Price the tool call, not the conversation. The conversation is what the user sees. The tool call is what you pay for.
The levers all sit before the build, while the vendor still wants your signature. Pin the units, cap the rates, and instrument the fan out in the pilot.
Require the MCP tool call unit and the completion pass through as named, capped rates in the order form. Salesforce set the platform direction in its newsroom, but the rates are yours to lock.
Measure the real fan out and completion tokens in a pilot before committing production budget. Read the Headless 360 licensing pillar for the full layer map.
An MCP tool call is a single invocation of one of the more than 60 Model Context Protocol tools that Headless 360 exposes to an external agent. Each call is a discrete unit of work, and it is the point where metering attaches for headless agent traffic.
The MCP tool call unit is not separately published as of July 2026, so treat it as an undisclosed line to pin. In practice a tool call that triggers an Agentforce action draws 20 Flex Credits, near 10 cents, plus any Data 360 processing behind it.
Model completions add a token cost on top of the action. When an agent calls Claude Sonnet or GPT 5, the tokens for the prompt and the response are billed as a completion pass through, which stacks on the Flex Credit draw for the action itself.
Build the burn model from three inputs: tool calls per day, credits per call, and completion tokens per call. Multiply through at peak volume, then convert credits at 500 dollars per 100,000 and add the completion pass through to get the daily cost.
The strongest lever is pinning the MCP tool call unit and the completion pass through as named, capped rates in the order form. If those two lines stay blank, the vendor sets them after your agents are live and your leverage is gone.
The line items to pin, the Flex Credit burn model, and the buyer side levers before you go agent first. Independent. Buyer side.
Join procurement and IT asset leaders who get our benchmarks on Salesforce, SAP, Oracle, Microsoft, and Workday AI pricing. No vendor spin.
