DigitalOcean launched Agent Droplets on 1 October 2026, introducing monthly subscriptions that combine the infrastructure used to run AI agents. The Pro plan costs $50 a month, while Team costs $200 a month.
The plans bring agent execution, hosted inference, memory, session storage, and tool access into one spending allowance. They build on DigitalOcean Managed Agents, which remains in public preview.
The launch addresses a practical problem for developers: an agent may need several services before it can complete useful work. Computing time, model tokens, storage, and external tools can otherwise involve separate charges and different billing units.
What Does DigitalOcean Bundle in Agent Droplets?
Agent Droplets combine services for executing an agent, accessing models, preserving working state, and connecting to tools. DigitalOcean calls the execution layer Harness Runtime, which runs agent workloads inside isolated microVMs, small virtual machines used as execution environments.
The runtime supports persistent sessions, allowing work to continue across pauses. DigitalOcean lists support for tools and frameworks including Claude Code, Codex CLI, OpenCode, CrewAI, and LangGraph. Developers can therefore bring familiar agent workflows into the managed environment.
The inference component supplies access to models hosted by DigitalOcean. Its Action Gateway connects agents to more than 16,000 tools, according to the company, including browser automation, web search, and web fetching. The gateway also handles permissions and credential access.
These components cover different parts of an agent’s workflow. The runtime executes code, the model produces responses or decisions, and the gateway lets the agent act on external systems.
Agent Droplets should consequently be evaluated as a managed agent service. They are distinct from the conventional production hosting needed to keep an agent-built website, database, or customer-facing application online.
How Do Agent Droplets Pricing and Usage Limits Work?
Agent Droplets provide a monthly monetary allowance with discounts on eligible resources.
| Plan | Monthly subscription | Discount on eligible usage |
|---|---|---|
| Pro | $50 | 15% |
| Team | $200 | 20% |
The subscription provides a spending allowance, not unlimited computing time or tokens. DigitalOcean’s launch announcement describes an option to stop at the plan price. However, its current billing documentation says there is no setting to stop at the subscription limit: usage continues at standard rates from other available balances.
The model choice also affects the bill. DigitalOcean-hosted open models receive the discount. Frontier models such as Claude and GPT are charged at standard rates and can still draw against the subscription balance.
DigitalOcean’s published examples of agent hours, tokens, and completed tasks are workload estimates rather than fixed included quotas. A longer session or a more expensive model can change how much work the same allowance covers.
Unused plan funds generally expire monthly. Conventional cloud services, including ordinary Droplets and managed databases, remain separately billed. Developers should therefore budget for the complete application, including any infrastructure outside the agent subscription.
HostScore’s Take
Agent Droplets make the agent infrastructure easier to purchase, but the hosting decision still starts with the workload. Running an agent, serving a large language model, and hosting the application an agent creates involve different requirements.
For a team building an agent that uses hosted models and external tools, an integrated platform can reduce the services it needs to connect and administer. This is the distinction explored in our best AI hosting platforms comparison: application hosting, managed AI services, and custom model infrastructure serve different needs.
Teams serving their own model weights face a separate infrastructure choice. GPU memory, inference software, concurrency, and sustained utilization become central, as discussed in our recent LLM hosting comparison.
Dedicated GPU and bare-metal hosting can fit teams that need control over that serving environment. Atlantic.Net, for example, offers dedicated GPU and bare-metal options for workloads including model training and inference. These provide an alternative route when the requirement is dedicated hardware and a customized software stack (see screenshot above). Operating that environment still requires model deployment, security, monitoring, and backups, with responsibilities depending on the management arrangement.
DigitalOcean’s public-preview status also matters when planning production dependencies. Teams should evaluate service limits, support terms, and recovery behaviour alongside the subscription price before relying on the platform for daily business operations.
