Until now an AI agent reading or writing Salesforce data through the API cost the same as any other integration: it used your API limits and nothing else. Salesforce has announced that this ends for agents. This guide explains what is being metered, what is not, what is still unknown and what is worth doing before the meter starts.
What is a Headless Platform Interaction?
A Headless Platform Interaction, HPI for short, is one successful call made by a registered AI agent to a production Salesforce org, either through an MCP server or through a direct API call. Each one will draw Flex Credits, the same prepaid pool used by Agentforce, and will show up in Digital Wallet, Salesforce's consumption dashboard.
"Headless" means without the Salesforce interface: an agent working on your data from outside, such as Claude through the hosted MCP server or an agent built on another platform. "Registered" refers to the agent identity Salesforce is introducing, so that each agent works under its own credentials instead of a person's login.
Which calls are metered and which are not?
The line is drawn on two questions: is the caller a registered agent, and is the target a production org. Everything else stays as it is today.
Registered AI agent calling a production org over MCP
Metered as HPI?: Yes, per successful call
Registered AI agent calling a production org through the API
Metered as HPI?: Yes, per successful call
Agent calls to sandboxes, scratch orgs or Developer Edition
Metered as HPI?: No
Traditional integrations, such as a scheduled ERP sync
Metered as HPI?: No: current pricing and limits stay
People working in Salesforce in the browser
Metered as HPI?: No
Failed calls
Metered as HPI?: No
How much will an HPI cost?
Nobody can say yet. The usage type is on the Flex Credits rate card under the Customer 360 Platform, but the number of credits per call is listed as TBA. Flex Credits cost $500 per 100,000, so the bill will be calls times credits per call times half a cent; until Salesforce publishes the middle term, any figure you see is a guess.
What you can know today is the first term: how many calls your agents make. That is the number to measure now.
When does metering start?
Salesforce announced the change in a Help article on 17 September 2026 and has promised 30 days' notice before adding the rate and starting to meter. Industry reports say Salesforce is aiming at November for agent registration, the related security controls and the billing model; no metering date has been published. Customers with agents already working through the API are expected to register them at renewal.
Does it apply to Agentforce or to Claude?
Agentforce's own actions already have their own meter, 20 Flex Credits per standard action, as explained in the Agentforce pricing guide. Whether an agent like Claude connected through the Salesforce in Claude plugin counts as a registered agent generating HPIs has not been spelled out publicly. Treat it as an open question and get your account executive's answer in writing for your own setup.
What should you do before the meter starts?
Make an inventory of every agent and MCP client that touches production, including the ones a single team connected on its own. Give each agent its own connected app with scoped permissions, and keep system integrations on dedicated integration users, so usage can be attributed to the right owner.
Measure call volumes per workflow while it is free, and set Digital Wallet alerts as soon as the usage type appears. Move work that does not need an agent, such as nightly bulk reads, to a classic integration that stays off the meter. At renewal, ask for usage visibility, protection on the HPI rate and written answers on borderline cases.
If you want a second pair of eyes, describe your agents and integrations: within one business day you get a written quote for an inventory of what will be metered and how to keep it under control.
Key points
- Nothing is billed yet: the per-call rate is TBA and Salesforce promises 30 days' notice.
- Only successful calls by registered agents to production orgs count. Sandboxes and classic integrations do not.
- List every agent and MCP client that touches production now, while it costs nothing to find out.
- Give each agent its own connected app and keep system integrations on separate users.
- Use your renewal to ask for usage visibility, rate protection and the answer on borderline cases in writing.
Related services
Written by

Founder
Founder of Cloud Atelier, an IT consulting studio based in Rome.