If you run AI coding agents - Codex CLI, Claude Code, OpenCode, or a framework like LangGraph - today you have two realistic choices: run them on your own machine or on a rented server. DigitalOcean just added a third path, and it’s a genuinely new category for a provider that made its name on flat-rate Droplets.
In September 2026, DigitalOcean launched Harness Runtime and Action Gateway under a new “Managed Agents” umbrella. Harness Runtime gives every agent session its own isolated microVM, billed per second. Action Gateway is a managed MCP endpoint that brokers access to 16,000+ tools.[1][3] Together they are an attempt to turn “run an agent somewhere and let it use real tools” into a consumption-billed cloud product, the way you already pay for compute or storage.
This is squarely in the dev-hosting niche this site covers - so let’s look at what it actually is, what it costs, how it stacks up against running agents on a plain VPS, and where the lock-in risk is real.
What Harness Runtime actually is
Harness Runtime is a sandbox runtime for AI coding agents. Each session runs in its own Firecracker microVM, with pause, resume, checkpoint, and fork built in.[5] You can run Codex CLI, Claude Code, OpenCode, or agents built with LangGraph or Hermes - or register your own OCI container image as a template (Bring-Your-Own-Template) instead of using one of the stock ones.[5][6]
The isolation is real, and it’s the interesting part for a hosting site. Isolation is enforced by the hypervisor, not by container namespaces: a process in one sandbox cannot observe or reach another session, another customer, or the host.[7] That is a different trust boundary than a long-running Droplet shared with unrelated jobs or containers on the same kernel. For agent work - where code runs untrusted input, spawns processes, and browses the web - that boundary matters more than it does for a normal web app.
A session is a microVM sized by an environment spec. You define the environment (which agent, which sandbox size, what permissions and tools), start a session from the Control Panel or with doctl, and either watch it in an attached terminal or let it run triggered by scheduled or webhook calls.[6]
What it costs
The headline number is $0.0535/hour - but that’s the price of the smallest sandbox at full allocation, not a flat rate.[1] Let’s be precise, because per-second cloud billing confuses people.
Compute is billed per second of active compute based on actual CPU consumption and the peak memory the sandbox used. When an agent is waiting and consuming no CPU, its CPU charge drops to zero.[1][2] The unit rates are:
- CPU: $0.044 per vCPU-hour (actual CPU consumed)
- Memory: $0.0095 per GB-hour (peak memory used)
- Session storage (volumes): $0.05 per GiB-month
- Public internet egress: $0.01 per GiB[1][2]
One honesty caveat straight from the source: true “actual CPU consumed” billing is coming soon. Until then you’re billed at 25% of the vCPUs allocated to your sandbox, whether the agent uses them or not - paused sessions incur no compute charge.[1][2] So the “zero when idle” promise is partially future-tense right now.
The sandbox sizes, at full allocation:
| Shape | vCPUs | Memory | Price at full allocation |
|---|---|---|---|
| XSmall | 1 | 1 GB | $0.0535/hr |
| Small | 2 | 2 GB | $0.107/hr |
| Medium (default) | 2 | 4 GB | $0.126/hr |
| Large | 4 | 8 GB | $0.252/hr |
| XLarge | 16 | 32 GB | $1.008/hr |
Actual charges follow consumption, so a session averaging less than full utilization costs less than the table.[1] Pausing a session stops compute charges while preserving the workspace and conversation history.[2] Kept snapshots and checkpoints continue to accrue storage charges while paused - $0.05/GiB-month for retained snapshots/checkpoints, and the same $0.05/GiB-month for custom BYOT template images.[1]
Two operational details matter before you adopt this:
- Prepaid balance required. Harness Runtime requires a positive prepaid account balance, and there’s no per-session or per-product spend limit. DigitalOcean itself warns that a looping or compromised agent using prepaid tools can drain that balance quickly.[1]
- No free tier, no per-app cap. This is consumption billing drawing down a shared balance, not a flat server fee. It is architecturally the opposite of a Droplet.
What Action Gateway does
Action Gateway is the tool-access layer. It gives agents governed access to 16,000+ tools through a single managed MCP endpoint or SDK, handling discovery, credential brokering, permissions, retries, and rate limits.[3][4] DigitalOcean-maintained tools include web search, web fetch, browser automation, and DigitalOcean infrastructure APIs, plus connectors for GitHub, HubSpot, Stripe, Snowflake and more - and you can extend the catalog with your own MCP servers.[4]
Pricing:
- Standard SaaS/MCP tool invocations: $0.10 per 1,000
- Web search: $10.00 per 1,000 tool calls
- Web fetch: $3.00 per 1,000 tool calls
- Tool search (
action_search): included at no additional cost - Code execution (
action_code): billed separately under Harness Runtime compute/memory[3]
Some tools - including Exa web search and web fetch - require a positive prepaid balance. When the balance hits zero, the gateway blocks calls to those tools until you add funds, but tools that don’t require prepayment are not blocked by the balance check.[3]
When you call Action Gateway from a Harness Runtime session, compute and storage bill under Harness Runtime while each tool call bills under Action Gateway.[1][3]
How this compares to running agents on a Droplet
This is where it gets relevant to anyone reading a hosting comparison on this site. We’ve covered the Docker vs raw VPS decision and the Droplet-vs-Linode-vs-Vultr tier, so the mental model is already here: a Droplet is a flat-rate, always-on server. It bills you whether the CPU is busy or not.
Harness Runtime is the inverse: it wants to bill only for active consumption, per second, drawing down a prepaid balance. The trade-off is not a pure win for either side:
- A Droplet is predictable. $6/month for the Basic 1 vCPU / 1 GiB plan whether you run one agent job or a hundred.[8] If your agents are lightly used or bursty-but-small, a $6 box and manual tooling may outperform per-second billing on cost.
- Harness Runtime is hands-off. No server to patch, no sandbox isolation to build yourself, pause/resume/checkpoint/fork as a managed primitive, and permission/approval controls are part of the product rather than something you bolt on.[6]
The honest framing: if you already have a VPS and a toolchain for running agents, the management overhead load that Harness removes may be small for you - and the flat box may be cheaper. If you’d rather not maintain agent infrastructure at all and you value isolation and lifecycle primitives being someone else’s problem, the consumption model is the point.
We’re not going to publish a dollar-vs-dollar comparison here, because an honest one requires measuring real agent workloads on real paid accounts - see how we test. Treat the numbers above as provider pricing structure, not a benchmark.
Why this is new for the small stack
For years the “managed AI infrastructure” story was tied to big clouds with big complexity - Kubernetes clusters, GPU fleets, platform engineering. DigitalOcean’s entry is notable because it targets the small stack that the rest of this site assumes: a solo developer or small team that wants managed primitives without a dedicated infra person.
The microVM-per-session isolation is the strongest argument. Untrusted-agent execution is exactly the case where container namespace isolation feels thin and a fresh sandbox per session is valuable. Getting that as a per-second-billed managed product, with checkpoints and forks as API primitives, is a materially different offering than “here’s a box, run whatever you like” - even if it’s a very cheap box.
The lock-in you should care about
Every managed product hides a lock-in cost, and this one has several worth naming:
- Prepaid-balance wallet lock. Both products draw down a shared prepaid balance tied to a DigitalOcean account. Your running costs live in DigitalOcean’s wallet model, not a normal postpaid bill.
- Catalog gravity. Action Gateway’s tool catalog, credential brokering, and policy layer are easiest to use via its managed MCP endpoint. Migrating that to a self-hosted MCP setup later is real work - it’s a managed answer to the MCP-server sprawl problem, and managed answers are sticky.
- Model/provider lock is weaker than it looks. Billing for models is explicit: DigitalOcean-hosted models pass through to Inference billing, and external provider keys (Anthropic, OpenAI) are billed directly by those providers.[1] So you’re not locked to DO’s models - but the sandbox + gateway orchestration is a DO-shaped dependency.
The practical mitigation is the same as with any PaaS: keep your environment spec, session scripts, and tool config in version control and out of a proprietary console, so the cognitive and config investment survives a move. The data and conversation state is in DigitalOcean’s sandboxes while a session runs, so treat session boundaries as the natural and only migration unit.
Who it’s for, who it’s not
It fits if: you run AI coding agents regularly and want managed isolation, lifecycle control (pause/checkpoint/fork), and governed tool access without building any of it; you’re comfortable with consumption billing and a prepaid balance; you want the agent-workload equivalent of “just use the managed thing.”
It fits if you’re not careful if: you prefer a flat monthly bill; your agent usage is tiny and a $6 entry Droplet plus some scripts covers it; you need strict per-vendor budgeting that a shared prepaid balance fights against; or you’d rather keep tool access self-hosted and fully on your own network.
The summary
DigitalOcean’s Managed Agents - Harness Runtime plus Action Gateway - is a real new category for a host that built its name on flat-rate servers: per-second consumption billing for AI agent execution in hypervisor-isolated Firecracker microVMs, plus a managed MCP gateway to 16,000+ tools.[1][3][5]
For a solo developer or small team already comfortable with DigitalOcean, it removes most of the infrastructure work from running trustworthy, well-instrumented agents. But it is consumption billing against a prepaid wallet, with real catalog and configuration gravity - so weigh the flat-rate predictability of a plain VPS against the hands-off isolation before you commit your agent workload to it. As with every managed primitive, keep the parts you can (specs, scripts, tool config) portable. If the appeal is containment without a cloud subscription - isolating agents on a machine you already manage - Microsoft’s Execution Containers (MXC) are the OS-level counterpoint to this whole category.
If that’s the workload you want to hand off, you can start an account at DigitalOcean - or head straight to the product page: DigitalOcean Managed Agents. Both links are affiliate, but they don’t change the verdict above. See our affiliate disclosure.
Sources
- DigitalOcean - Harness Runtime Pricing (sandbox sizes, compute/memory/storage/egress rates, prepaid balance, BYOT). https://www.digitalocean.com/pricing/harness-runtime - checked 2026-09-29
- DigitalOcean Docs - Harness Runtime Pricing (per-second active compute, peak-memory billing, 25%-of-vCPUs interim rate, pause semantics). https://docs.digitalocean.com/products/managed-agents/agent-harness-runtime/details/pricing/ - checked 2026-09-29
- DigitalOcean - Action Gateway Pricing ($0.10/1k invocations, web search $10/1k, web fetch $3/1k, prepayment). https://digitalocean.com/pricing/action-gateway - checked 2026-09-29
- DigitalOcean Docs - Action Gateway (managed MCP endpoint, 16,000+ tools, connectors). https://docs.digitalocean.com/products/managed-agents/action-gateway/ - checked 2026-09-29
- DigitalOcean Changelog - “Harness Runtime is now in Public Preview” (Firecracker microVM, pause/resume/checkpoint/fork, supported agents, BYOT). https://ideas.digitalocean.com/changelog/harness-runtime-is-now-in-public-preview - checked 2026-09-29
- DigitalOcean Docs - Harness Runtime Overview (environment spec, doctl, sandbox, adapters, egress allowlists). https://docs.digitalocean.com/products/managed-agents/agent-harness-runtime/concepts/overview/ - checked 2026-09-29
- DigitalOcean Docs - Environments and Sandboxes (hypervisor-level isolation vs container namespaces). https://docs.digitalocean.com/products/managed-agents/agent-harness-runtime/concepts/environments/ - checked 2026-09-29
- DigitalOcean - Droplet pricing (Basic 1 vCPU / 1 GiB - $6/mo). https://www.digitalocean.com/pricing/droplets - checked 2026-10-03