Most SaaS pricing assumes human users. A seat license makes sense when a human being sits down at a computer, logs in, and does work for a predictable number of hours per month. AI agents do not work that way. They might be idle for 23 hours and then process a thousand tasks in one hour. They might run on behalf of a dozen users simultaneously, or they might run on behalf of nobody \u2014 autonomously acting on a scheduled trigger.
Seat-based pricing punishes this usage pattern. You either overpay for idle capacity or you scramble to upgrade your plan when an agent has an unexpectedly busy week. Neither is acceptable for a business trying to build a reliable cost structure around AI.
What ARC Credits Are
ARC (Agent Resource Credits) is CloudClaw's unit of platform consumption. Every action an agent takes on the platform \u2014 an LLM call, a connector invocation, a stored memory write, a tool execution \u2014 costs a defined number of ARC credits. Credits are purchased in bundles or earned through the referral and marketplace programs.
The model has three properties we care about deeply:
- Proportionality. You pay for what your agent actually does, not for having deployed it. A dormant agent costs nothing.
- Predictability. Each action type has a published credit cost. You can model your monthly spend before you write a line of code.
- Transferability. Credits can be allocated across agents, across tenants, and (through the marketplace) exchanged between platform participants.
The Marketplace Dimension
The credit economy becomes more interesting when you introduce the marketplace. Platform participants who build high-quality agents and publish them to the CloudClaw Marketplace earn ARC credits every time another business deploys their agent. Those credits can be redeployed to fund the builder's own agent operations.
This creates a flywheel: good agents generate credits, credits fund more development, better agents attract more deployments. It is the same logic that made app stores transformative, but applied to the AI agent layer of the stack.
Designing for Marginal Cost
We made a deliberate infrastructure choice to run CloudClaw on Cloudflare's global edge network. The marginal cost of an additional agent execution is extremely low at this layer \u2014 which is why we can offer usage-based pricing without a large fixed cost floor. When your agent's marginal cost is low, your pricing can be honest rather than defensively padded.
If you are evaluating AI agent platforms and you are comparing flat-fee seats to ARC credits, ask yourself: what does my usage pattern actually look like? If it is bursty, seasonal, or hard to predict, usage-based wins. See the full pricing breakdown and run the numbers for your use case.