Convergence.ai licence cost: Proxy Lite, Agentforce and the screen-driving agent

The licence trap with Convergence is not in Salesforce's price book at all. It is in a code repository. A developer finds a small open model that can drive a browser, proves it on a laptop, and a proposal arrives to run it against production systems. Whether that is allowed depends on a licence file most people do not open.

#Three things people mean by Convergence

The acquisition itself is simple.

"Salesforce completed its acquisition of Convergence on June 11, 2025."

The announcement described Convergence as an AI agent company whose systems can navigate dynamic interfaces, adapting to pop-ups, errors and interface changes, and said its team and technology were expected to play a central role in advancing Agentforce. That description fits three different things a buyer may be holding, and each has a different licence.

The first is the open model. Convergence published Proxy Lite, which the repository describes with the sentence "A mini, open-weights, version of our Proxy assistant." The second is whatever hosted product existed before the acquisition. Its website returned a forbidden response when we tried it on 13 October 2026, so we could not read any terms and we will not describe them. The third is what Salesforce now sells that serves the same purpose. That is the part with a price book, and it is where most of this page lives.

#The open model and its licence

The repository's licence file is short, and it settles the question for most commercial teams. It states: "This work is licensed under the Creative Commons Attribution-NonCommercial 4.0 International License." A noncommercial licence permits use that is not primarily directed toward commercial advantage or monetary compensation, and an agent that processes your company's invoices, orders or customer records is hard to describe as noncommercial. The licence text is public, and the sensible reading for a business is that you need separate permission from the rights holder before using the model in operations.

The practical consequences:

• Using Proxy Lite for a hackathon, a classroom or personal research is the use the licence plainly contemplates.
• Using it inside a company workflow, even a free internal one, is a legal question to put to counsel before the first run.
• Fine-tuning it on your own screens and redistributing the result inherits the same restriction.
• Downloading weights is not buying a licence. There is no price you can pay on the repository to change the terms.

This is the cheapest line in the whole subject to get wrong, because it costs nothing until it costs a great deal. If your proof of concept already runs on it, record that and plan the commercial route before anyone depends on the output.

#What Salesforce sells for the same job

Salesforce prices Agentforce in more than one way, and the Agentforce pricing page read on 13 October 2026 shows three that matter for an agent that takes actions.

Flex Credits are the consumption unit. The page defines them as "a new unit of payment for Agentforce that align cost to the business value your AI agents create", and lists US$500 per 100,000 credits. An agent action draws 20 credits, so each action costs US$0.10, and a voice action draws 30.

The employee add-on is the seat route. Agentforce add-ons for Sales, Service and Field Service and Operations were listed at US$125 per user per month, with unmetered Agentforce use for employees. An Industries add-on was listed at US$150.

Max editions are the bundle. They were listed from US$550 per user per month, with the add-on included and an allocation of 2.75 million Flex Credits per org per year.

The rules around credits matter as much as the rates. "Unused Flex Credits do not roll over into subsequent subscription terms." Credits and conversation pricing cannot be mixed in one org. Pay-as-you-go, pre-commit and pre-purchase buying models exist, and on pre-commit a true-up applies at term end if usage falls below your commitment. Overage is billed at your contracted rate, monthly in arrears.

#The agent that learns a browser

The closest thing in Salesforce's current range to a screen-driving agent is on the Agentforce Operations page. That page says: "Bridget works anywhere, handling any browser-based app, including legacy systems with no APIs." It adds that the agent learns how an application works in about an hour, with no code, APIs or RPA scripts to maintain, and that it relearns when a user interface changes.

We cannot tell you that this agent is built on Convergence technology. The acquisition announcement does not say so and the product page does not mention Convergence. What I can tell you is that it is the product to ask about if your requirement is an agent that works through an interface because no API exists, and that no price appears on that page. Treat it as quote-only and ask what it meters. Candidates include the number of systems learned, the number of process runs, the number of users, and consumption in Flex Credits. The page does not say which, so the answer belongs in your written quote.

#When an API makes the licence question disappear

Screen-driving agents are what you reach for when the system cannot be reached any other way. If the target has an API, the cleaner licence is almost always the API. An API call is governed by the vendor's published developer terms, it is stable when the interface changes, and it does not depend on a model deciding where a button is.

Before accepting a screen-driving design, run a short test of the target system.

1. Does it have a documented API that covers the transaction you need? If yes, use it and price the integration instead.
2. If not, do its terms permit automated access to its interface? Many commercial terms restrict bots and scraping, and a breach is your contractual risk, not Salesforce's.
3. Who is the user? An agent that logs in needs a licensed user in the target system, and that may be a named seat you must buy.
4. What does a failure look like? A screen change that breaks the agent silently is worse than a failed API call, so ask how failures are surfaced.

The earlier guide on Convergence and Agentforce walks through that API-first test in more detail.

#Worked arithmetic with your own volume

You can bound the credit cost of a screen-driving workload before you buy, using only published rates and a volume you choose. If a process run needs a handful of agent actions, multiply the actions by 20 credits and by US$0.005 a credit, which is the published rate of US$500 per 100,000. A run of ten actions is 200 credits, or US$1. Multiply by monthly runs. At 5,000 runs a month that is 1,000,000 credits and US$5,000, and you can set it beside a US$125 seat times your user count to see which route is cheaper. The figures here are an illustration of the method, not a forecast, and your action count per run is the number that matters, so measure it in a pilot.

#The licence the agent logs in with

An agent that works through a screen has to sign in, and signing in needs a user. That user is licensed by the system being driven, not by Salesforce, and it is the cost most often missing from a proposal. Three questions settle it.

Does the target system sell a named seat, an integration user, or a service account, and what does each cost? A named seat for a bot can be dearer than the Salesforce cost of the agent itself.

Do its terms allow an automated login? Some vendors permit service accounts for integration but forbid scripted use of the interface, and some require approval. A breach is your contractual exposure, and a vendor that notices can suspend the account the process depends on.

Is the data entered by the agent attributable? Audit trails matter when a regulated record is changed. Ask whether the target system will show that a change came from an agent identity, and whether your auditors will accept that.

None of those answers appear in any Salesforce price list, which is why they belong in the business case from the first page.

#Reading the three licences together

Put the three possible routes side by side before choosing, because they differ in kind and not just in price.

#What a proof of concept should record

A proof of concept is where a licence problem is either caught or buried. Whatever the route, keep a short written record of the following, so that the production decision is made on facts.

1. Which model, version and licence the agent ran on, with a copy of the licence text as it stood on the day.
2. Which systems it logged into, as which user, and whether the terms for each were checked.
3. How many agent actions each completed process used, because that is the number that turns into credits.
4. How often the agent failed, how the failure was surfaced, and who was told.
5. Which data left your boundary, and where the model that processed it was hosted.

Item three is the one that pays back most. A rough estimate of actions per process run can be wrong by a large factor in either direction, and the whole credit budget hangs on it. Run a hundred real cases, count the actions, and multiply by the published 20 credits an action.

#Questions to put in the quote

Ask whether the quote is credits, seats or a bundle, and what the rate is above the allowance. Ask whether voice or other action types draw credits at a higher rate, as the pricing page shows for voice at 30 credits. Ask what happens to credits at the end of the term. Ask which systems the browser-learning agent is licensed to learn, and whether a new system is an additional charge. Ask for those answers to be written into the order form, since a conversation with a seller is not a contract.

#A decision path for a buyer

If you are weighing these routes, a short sequence avoids most of the expensive mistakes. First, check whether the target has an API for the transaction, and if it does, price the integration and stop. Second, if it has no API, read the target's terms on automated access and ask the vendor for permission in writing. Third, decide whether the agent will act for employees or for customers, because that decides whether the seat route or the credit route fits. Fourth, run a pilot that counts actions per process, and size credits from that count. Fifth, ask Salesforce for the quote on the browser-learning agent only if the first four steps leave a real gap, since it is the one route with no published price.

That order matters because each step is cheaper than the next. A licence question settled by an API is free, a terms question settled by a conversation is cheap, and a credit estimate settled by a pilot is modest. The expensive mistakes come from starting at the bottom with a clever agent and discovering the licence problems in production.

#Where the neighbouring questions are answered

The acquisition date and what the announcement said are on what happened to Convergence. The shape of a sound integration, with an object and a failure path, is in integrating Convergence with Salesforce, and the decision about whether to let an agent operate a given system is in should I migrate from Convergence.

#Lines to separate on the order form

Keep these as separate lines in your own budget, and ask your account executive to confirm in writing which of them the quote includes.