DigitalOcean Managed Agents
DigitalOcean · DigitalOcean Managed Agents: Managed Cloud Runtime and Tool Gateway for AI Agents
DigitalOcean Managed Agents is a cloud platform for running existing AI agents, such as Claude Code, Codex CLI, OpenCode, Hermes, or LangGraph apps, in isolated Firecracker microVMs that can pause, resume, and fork. Its Action Gateway gives those agents governed access to 16,000+ tools through one MCP endpoint, with credentials kept out of the model and sandbox. It suits developers and small teams who want agents running durably in the cloud without building the sandbox, session, and tool plumbing themselves.
Best for
Developers and small teams who already use Claude Code, Codex, OpenCode, Hermes, or LangGraph and want those agents running durably in the cloud, with pause, resume, and fork, governed tool access, and usage-based billing in one place
Not ideal for
Teams that need self-hosting, private VPC access, or regions outside RIC1, buyers who need configurable spend limits, and anyone looking for a ready-made agent rather than infrastructure to run one
Who it's for
Engineering teams and individual developers running coding or general-purpose agents, and companies building agent products that need per-tenant sandboxes and tool governance
Capabilities
- Harness Runtime runs each agent session in a dedicated Firecracker microVM with its own compute, filesystem, and built-in tools such as Chromium
- Adapters for Claude Code, Codex CLI, OpenCode, Hermes, and LangGraph applications, a codex-agentapi adapter for OpenAI Agents API sessions, and a custom adapter for your own OCI image
- Sandboxes can also be used on their own to execute code for an agent or app that runs elsewhere, such as generated Python, tests, or file processing
- Bring-Your-Own-Template registers a team-owned OCI image as a reusable sandbox template
- Pause and resume with memory, filesystem, and conversation history preserved, plus automatic pause on idle (15 minutes by default) so compute and memory charges stop
- Checkpoints capture a session's machine state so you can fork a new session from it or roll back, with support varying by adapter
- YAML environment specs cover adapter, sandbox size, secrets, skills, permissions, egress allowlist, and tools, and saved Environment Configs start any number of sessions
- Permission policies set allow, ask, or deny defaults with rules for shell commands and file paths, and require human approval for sensitive actions
- Scheduled and webhook triggers run agents unattended and can deliver output to email or Slack
- Exec, file upload and download (up to 50 GiB per file), and authenticated port forwarding through doctl to reach apps running in the sandbox
- Action Gateway exposes 16,000+ tools across 500+ providers, such as GitHub, HubSpot, Stripe, Snowflake, PagerDuty, Box, Supabase, and Exa, through one managed MCP endpoint, plus your own MCP servers
- Credential brokering for static API keys, shared OAuth apps, or per-user OAuth, with credentials resolved at execution time and kept out of the model, prompt, and sandbox
- Tool policies that allow, require approval for, or deny calls, versioned Toolbelts, tool search, and built-in rate-limit handling, retries, and backoff
- Structured event logs of tool calls, model requests, and file operations, with token usage and approval activity per session
Limitations
- Managed Agents and the Agent Droplets plans are in public preview (you opt in to the preview terms), and Harness Runtime is currently available only in DigitalOcean's RIC1 region
- Active CPU billing is not live yet. Until it ships, CPU is billed at 25% of the vCPUs allocated to the sandbox rather than actual consumption
- There is no configurable per-session or per-product spend limit. When an Agent Droplets balance runs out, usage continues at standard rates from promotional credits, the Inference and Agents balance, and the account prepayment balance, and sessions pause only when all of them are empty. A looping agent using prepaid tools can drain these balances quickly
- Agent Droplets discounts do not apply to third-party models such as Anthropic Claude or OpenAI GPT. The documented Claude Code and Codex setups use your own Anthropic or OpenAI API key, billed by that provider
- Harness Runtime sandboxes cannot connect to private VPC resources, VPC peering is not supported, and model access keys scoped to a VPC network cannot be used
- Sandbox egress is unrestricted unless the environment spec sets an allowlist
- Concurrent sessions are limited by your account's resource tier, up to 100 per team unless you request an increase, and each session has a 744-hour monthly active compute cap
- The codex-agentapi adapter skips DigitalOcean's event stream, permission policies, approvals, checkpoints, forks, and Action Gateway tools
- CrewAI is named on the product page and in the Agent Droplets announcement, but the documented adapter list has no CrewAI adapter. Agents without an adapter run through the custom container image adapter, with a less structured event stream
- No self-hosted deployment is documented. Agents run only on DigitalOcean's cloud
Use cases
- Starting a long Claude Code refactor in a cloud sandbox, closing the laptop, and reattaching later from another device
- Firing an unattended first-pass pull request review from a GitHub webhook and posting the result to Slack
- Running a scheduled research or reporting agent that writes its output to a persistent workspace
- Forking a checkpointed session to try several approaches in parallel and keeping the one that works
- Giving an existing MCP client governed access to GitHub, Stripe, or Snowflake through Action Gateway without handing it raw credentials
- Starting an isolated session per customer from one saved environment for a multi-tenant agent product
Our take
Managed Agents is a good fit for the developer who already has a working agent setup and is tired of it living on a laptop. The useful parts are the session lifecycle, which lets long tasks pause, resume, and branch without rebuilding context, and Action Gateway's credential brokering, which can be used from any MCP client. The pricing pitch needs a careful read. Active CPU billing is not live yet, there is no configurable spend limit, and the Agent Droplets discount skips Claude and GPT, so the savings are largest for teams willing to run DigitalOcean-hosted open models. As a single-region public preview without private VPC access, it is better suited to development workloads and new agent products than to replacing an established production stack today.
Who should use it
Developers and small teams running Claude Code, Codex, OpenCode, Hermes, or LangGraph who want cloud sessions they can pause, resume, and fork, startups building agent products that need per-tenant sandboxes and governed tool access, and teams willing to use DigitalOcean-hosted open models to get the most from the plan discounts.
Who should skip it
Teams that need self-hosting, private VPC connectivity, or data residency outside RIC1, organizations that require configurable spend limits, and buyers who mainly use Claude or GPT and expect the plan discount to cover model costs.
Strengths
- Runs agents you already use instead of a proprietary harness
- Pause, resume, checkpoint, and fork make long or branching agent work practical
- Action Gateway keeps credentials out of the model and sandbox and can be used from any MCP client
- Published usage rates, with optional plans that discount compute, storage, hosted inference, and tools on one bill
- No seat charges, and unlimited agents on Agent Droplets plans, subject to fair use and concurrency limits
Weaknesses
- Public preview, with Harness Runtime in a single region
- No configurable spend limit, and active CPU billing is not live yet
- Plan discounts exclude Claude, GPT, and other third-party models
- No private VPC access from sandboxes and no self-hosted option
DigitalOcean Managed Agents pricing
Pay as you go
Usage-based
- CPU $0.044 per vCPU-hour (billed at 25% of allocated vCPUs until active CPU billing launches)
- Memory $0.0095 per GB-hour of peak usage
- Session, snapshot, and custom template storage $0.05 per GiB-month
- Internet egress $0.01 per GiB
- Action Gateway $0.10 per 1,000 tool calls, with Exa web search at $10.10 and web fetch at $3.10 per 1,000 calls
Free Trial
$5 credit
- New DigitalOcean customers only
- Applies to Harness Runtime, Action Gateway, and Serverless Inference
Agent Droplet Pro
$50
Billed monthly
- 15% off eligible Managed Agents usage
- Covers $58.82 of standard-rate usage
- Unlimited agents and seats, subject to fair use
- SSO, MFA, role-based access, and audit logs included
Agent Droplet Team
$200
Billed monthly
- 20% off eligible Managed Agents usage
- Covers $250 of standard-rate usage
- Unlimited agents and seats, subject to fair use
- SSO, MFA, role-based access, and audit logs included
- Upgrade from Pro takes effect immediately
Free tier limits: New customers receive a $5 promotional credit for Managed Agents usage. Existing customers do not receive it. There is no ongoing free tier.
Note: Agent Droplets plans are in public preview and discount Harness Runtime compute, memory, storage, egress, DigitalOcean-hosted inference (such as Kimi K3 and GLM-5.3), and Action Gateway tools. Third-party models such as Anthropic Claude and OpenAI GPT are not discounted, and provider keys you bring are billed by that provider. Plan fees are charged in advance, are not refunded, and unused subscription balance expires monthly except in the first month and after an upgrade. When the plan balance runs out, usage continues at standard rates from promotional credits, the Inference and Agents balance, or the account prepayment balance, with no setting to stop it, and sessions pause only when all balances are empty. A Medium sandbox (2 vCPU, 4 GB) is listed at $0.126 an hour at full allocation, or about $0.06 an hour for CPU and memory while CPU is billed at 25% of allocated vCPUs.
Technical specs
API pricing
Usage-based: $0.044 per vCPU-hour, $0.0095 per GB-hour of peak memory, $0.05 per GiB-month of storage, $0.01 per GiB of egress, and $0.10 per 1,000 Action Gateway tool calls; DigitalOcean-hosted inference billed at Inference rates
Where DigitalOcean Managed Agents excels
Long-running coding sessions that survive a closed laptop
A Claude Code or Codex session keeps working in its microVM, pauses on idle so compute charges stop, and can be reattached from another device with the workspace and conversation intact.
Unattended agents on schedules and webhooks
Triggers can start an agent from a GitHub webhook or on a schedule, for example a first-pass PR review or a weekly research brief, and deliver the output to Slack or email.
Agent-powered products with per-customer isolation
One saved environment can start an isolated session per tenant, while Action Gateway brokers each end user's OAuth so credentials never reach the model or the sandbox.
DigitalOcean Managed Agents vs. competitors
DigitalOcean Managed Agents vs. NVIDIA OpenShell
NVIDIA OpenShell is open-source (Apache 2.0) software you run yourself on Docker, Podman, Kubernetes, or microVMs, with kernel-level YAML policies, credential placeholders, and a policy prover that checks new access before approval. DigitalOcean Managed Agents is a hosted service: DigitalOcean runs the microVMs and adds pause, resume, checkpoint, and fork, a 16,000-tool MCP gateway, and usage-based billing. OpenShell fits teams that need agents on their own infrastructure with verifiable policy, and Managed Agents fits teams that would rather not operate the sandbox layer at all.
DigitalOcean Managed Agents vs. Replicas
Replicas is built around the software team's workflow: tasks arrive from Slack, Linear, GitHub, or GitLab, each runs a coding agent in its own VM, and the resulting pull requests follow up on CI failures and reviews, priced per seat plus per-minute automation. DigitalOcean Managed Agents is a lower-level runtime: you define environments, start sessions from the CLI, API, schedules, or webhooks, and can run general-purpose agents such as Hermes or LangGraph apps as well as coding agents, billed by resource usage. Replicas fits teams that want a ready coding-agent workflow, and Managed Agents fits teams building their own agent workflows or products.
Frequently asked questions
What is DigitalOcean Managed Agents?
It is DigitalOcean's platform for running AI agents in the cloud. Harness Runtime runs agent sessions in isolated Firecracker microVMs with pause, resume, checkpoint, and fork, and Action Gateway gives agents governed access to 16,000+ tools through a single managed MCP endpoint. It does not include an agent of its own. You bring a supported harness or your own container image.
Which agents can it run?
The documented adapters are claude-code, codex (Codex CLI), opencode, hermes, langgraph (a LangGraph application), codex-agentapi (an OpenAI Agents API session using a DigitalOcean sandbox), and custom, which runs a container image you supply.
How much does DigitalOcean Managed Agents cost?
Harness Runtime charges $0.044 per vCPU-hour and $0.0095 per GB-hour of peak memory, $0.05 per GiB-month for session, snapshot, and template storage, and $0.01 per GiB of internet egress. Action Gateway charges $0.10 per 1,000 tool calls, with paid tools such as Exa web search priced separately. A Medium sandbox (2 vCPU, 4 GB) is listed at $0.126 an hour at full allocation. Until active CPU billing launches, CPU is billed at 25% of allocated vCPUs, which puts a running Medium sandbox at about $0.06 an hour for CPU and memory. Paused sessions incur no compute charges.
What are Agent Droplets?
Agent Droplets are optional monthly plans, in public preview, that prepay Managed Agents usage at a discount. Pro costs $50 a month and covers $58.82 of standard-rate usage (15% off). Team costs $200 and covers $250 (20% off). They do not add compute, models, or tools of their own, unused balance generally expires each month, and usage beyond the plan continues at standard rates.
Does an Agent Droplets plan cover Claude or GPT usage?
No. The discount applies to DigitalOcean-hosted models such as Kimi K3 and GLM-5.3, not to third-party models such as Anthropic Claude or OpenAI GPT, even through DigitalOcean Serverless Inference. If you use your own provider API key, as the documented Claude Code and Codex setups do, that provider bills you directly.
Is there a free trial?
New DigitalOcean customers get a $5 promotional credit for Managed Agents usage. Existing customers do not receive the credit and can subscribe to a plan or add funds instead.
Can I set a spending limit?
Not directly. DigitalOcean documents no per-session or per-product spend limit, and no setting stops usage when an Agent Droplets balance reaches $0. Harness Runtime runs on prepaid funds, drawing from the plan balance, then promotional credits, the Inference and Agents balance, and the account prepayment balance, and running sessions pause at the next safe stopping point only when all of them are empty. Running agents in a dedicated standalone team (one outside an organization), which keeps its own prepaid balances, separates agent spend from your other DigitalOcean usage.
Can I use Action Gateway without Harness Runtime?
Yes. Action Gateway works with any MCP-compatible client: you add its endpoint to the client's MCP configuration. It also has Python and TypeScript SDKs, and its quickstart connects it to Claude Code or Codex.
DigitalOcean Managed Agents vs NVIDIA OpenShell: what is the difference?
Both run existing agents such as Claude Code and Codex inside an isolated boundary and keep real credentials away from the agent. OpenShell is open-source software you install and operate on your own Docker, Podman, Kubernetes, or microVM hosts. Managed Agents is a hosted DigitalOcean service that adds session pause, resume, and fork, a tool catalog, and usage-based billing, but runs only on DigitalOcean.
Integrations & fit
Alternatives to consider
NVIDIA
Developers and platform or security teams who run coding or autonomous agents with real file, network, and credential access and want enforceable, reviewable limits outside the agent itself
Free
Replicas
Engineering teams already using Claude Code, Codex, or Cursor who want those agents running in configured cloud VMs, triggered from Slack, Linear, GitHub, GitLab, or automations, with PRs that follow up on CI failures and reviews
PaidAbout DigitalOcean Managed Agents
Managed Agents is not an agent. It is where you run one. It has two services that work together or separately. Harness Runtime runs each session in a dedicated Firecracker microVM with its own compute, filesystem, and built-in tools such as Chromium. You describe an environment in a YAML spec (or the Control Panel): the adapter (claude-code, codex, opencode, hermes, langgraph, codex-agentapi, or a custom OCI image), sandbox size, secrets, skills, permission policy, egress allowlist, and tools. Sessions keep running when your laptop closes, pause on idle to stop compute and memory charges, resume with files, processes, and conversation history intact, and can be checkpointed, forked, or rolled back, depending on the adapter. Scheduled and webhook triggers run agents unattended, and permission policies can allow, deny, or require approval for shell commands and file paths. Action Gateway is a managed MCP endpoint covering 16,000+ tools across 500+ providers, including GitHub, HubSpot, Stripe, Snowflake, PagerDuty, and Exa, plus your own MCP servers. It searches for relevant tools instead of loading the catalog into context, brokers API keys and OAuth at execution time, and applies per-tool policies and approvals. It also works with any MCP client outside Harness Runtime. Billing is usage-based: $0.044 per vCPU-hour and $0.0095 per GB-hour, plus storage, egress, and $0.10 per 1,000 tool calls. Optional Agent Droplets plans, launched October 1, 2026 ($50 Pro, $200 Team), prepay usage at a 15% or 20% discount, and new customers get a $5 credit. The tradeoffs come from the preview stage. Harness Runtime runs only in the RIC1 region, active CPU billing is still coming (CPU is billed at 25% of allocated vCPUs until then), there is no configurable spend limit, sandboxes cannot reach private VPC resources, and plan discounts do not cover third-party models such as Claude or GPT, which Claude Code and Codex use through your own API keys.
Updates from DigitalOcean Managed Agents
DigitalOcean introduced Agent Droplets, monthly plans in public preview that prepay Managed Agents usage at a discount: Pro at $50 a month (15% off) and Team at $200 a month (20% off). The discount covers compute, memory, storage, DigitalOcean-hosted inference, and Action Gateway tools, but not third-party models such as Claude or GPT. An Inference and Agents prepaid balance launched alongside it.
After a private preview, DigitalOcean opened Managed Agents to all users in public preview, made up of Harness Runtime for microVM agent sessions and Action Gateway for governed access to 16,000+ tools through one MCP endpoint. Teams can also register custom OCI images as sandbox templates.
Are you the founder? Claim this listing →