Back to vendors
D

DigitalOcean

Also known as: digitalocean.com, DigitalOcean AI Platform, DigitalOcean Gradient AI Platform, DigitalOcean Agent Platform, DigitalOcean Managed Agents, DigitalOcean Gradient ADK

Visit site
Entry priceAgent creation free; pay per token, such as GPT-5 mini at $0.25 per million input and $2 per million output tokensFull pricing detail

DigitalOcean is a developer cloud whose AI Platform lets teams build and host agents next to their apps and data. Developers pick a model from OpenAI, Anthropic or open source families, attach knowledge bases on OpenSearch, route between agents and functions, add guardrails, and test changes with agent evaluations, then call each agent through its own API endpoint or a chatbot embed. A code first Agent Development Kit and Managed Agents, with microVM sandboxes and a tool gateway, are in public preview.

DigitalOcean sells cloud infrastructure aimed at developers, startups and smaller software companies, from Droplet virtual machines, Kubernetes and App Platform to managed databases, Spaces object storage, serverless Functions and GPU servers. DigitalOcean Holdings, Inc. is based in Broomfield, Colorado and trades on the New York Stock Exchange, and it now sells the whole estate as an AI native cloud for inference and agents.

AI customer annual recurring revenue reached $234 million in the second quarter of 2026, up 212% on a year earlier, and Hippocratic AI, Ideogram, Workato and Fal.ai build on its AI products. Those products sit on the same account, network and invoice, from serverless and dedicated model inference to an inference router that picks a model per request.

The Agent Platform builds managed agents from the control panel, the CLI or the API. Each agent gets a name, instructions, a foundation model chosen from OpenAI, Anthropic, Mistral, Llama, DeepSeek and other families, a project and a datacenter region, and it can be grouped with others in a workspace. Knowledge bases ground answers in files from Spaces, local uploads, seed or sitemap URLs, Dropbox and Amazon S3, indexed into OpenSearch with a choice of embedding model, chunking method, semantic, keyword or hybrid retrieval and reranking, with auto indexing when sources change.

Function routes let an agent call code for live or outside data, agent routes pass a query to the right agent in a multi agent setup, and guardrails scan inputs and outputs for sensitive data, jailbreak attempts and harmful content. Agent Evaluations act as a test suite: teams define prompts and metrics for quality, safety, context use, tool use and goal completion, run them after a change on a set schedule and read pass and fail rates. Traces show inputs, model outputs and function call results, and each agent ships with an API endpoint, access keys, a playground, versions and a chatbot snippet.

Two newer products are in public preview, which DigitalOcean does not recommend for sensitive workloads: the Agent Development Kit, a Python SDK and CLI that deploys agent code with automatic tracing, and Managed Agents, which runs harnesses such as Claude Code, Codex and OpenCode in microVM sandboxes with Chromium and reaches more than 16,000 tools through an Action Gateway MCP endpoint, with tool call approvals in the Cuttlefish iOS app.

Agent creation is free, and agents bill model tokens at serverless rates, such as $0.25 per million input tokens and $2 per million output tokens for GPT-5 mini, or $3 and $15 for Claude Sonnet 4.6, with knowledge base embedding, reranking, guardrails and functions billed separately. Customers can bring their own OpenAI or Anthropic key. DigitalOcean holds SOC 2 and SOC 3 reports, offers predefined and custom team roles, and runs the Agent Platform in Toronto, Atlanta and Richmond.

It fits developers and small software teams that already host on DigitalOcean and want agents, retrieval, evaluation and model choice on one bill. It fits less well for enterprises that need agent hosting outside North America, a production grade sandbox runtime today, or agents that wake on schedules and events without their own code.

Vendor details

Canonical URL

https://www.digitalocean.com

Category

Agent infrastructure

Subcategory

Developer cloud with a managed agent platform, model inference, knowledge bases, agent evaluations and sandboxed agent runtimes

Funding status

Public company. DigitalOcean Holdings, Inc., headquartered in Broomfield, Colorado, trades on the New York Stock Exchange (DOCN); DigitalOcean, LLC is its operating company.

Company status

independent

Use cases & customers

Primary use cases

hosted agents with retrieval over company documentsmulti agent routing across specialist agentsagent evaluation suites run after every changeserverless model inference across commercial and open modelssandboxed runs of coding agent harnesses in preview

Target customers

developers and startups building AI applicationssmall and midsize software companies hosting on DigitalOceanteams that want inference, retrieval and agents on one cloud bill

Deployment options

Managed agents in TOR1, ATL1 or RIC1Serverless and dedicated inferenceMicroVM sandboxes (Managed Agents, public preview)Agent code deployed with the ADK (public preview)

Integrations

Agents call code through function routes and hand off to other agents through agent routes. Knowledge bases ingest DigitalOcean Spaces, local files, website and sitemap URLs, Dropbox and Amazon S3, with a Knowledge Bases MCP server for retrieval. Models come from OpenAI, Anthropic, Mistral, Llama and DeepSeek families, with bring your own key for OpenAI and Anthropic. In preview, Managed Agents' Action Gateway reaches more than 16,000 tools through one MCP endpoint, with connectors for GitHub, HubSpot, Stripe and Snowflake. Agents are reachable through their own API endpoints, the DigitalOcean API and CLI, and a chatbot embed.

In practice

A SaaS startup builds a support agent on Claude Haiku 4.5, attaches a knowledge base indexed from its docs sitemap with hybrid search, adds a sensitive data guardrail and embeds the chatbot snippet in its app.

An ecommerce team runs a router agent that sends billing questions to one agent and product questions to another, and a function route looks up live order status from its own API before the agent replies.

Before switching a production agent from GPT-4o to GPT-5 mini, an engineer reruns the team's evaluation set and compares pass rates on goal completion and context use, then promotes the new version once the scores hold.

Agentic Index coverage score

12.0 / 14 capabilities · 86%

Integrations & Tool Calling Full

Agents reach outside systems through function routes, which hand a request to code the team writes and bring live or external data back into the answer, and through function calling to third party tools. On the inference side, a server side MCP tool lets models call remote MCP servers. In public preview, Managed Agents adds an Action Gateway that exposes more than 16,000 tools through one managed MCP endpoint, with connectors for GitHub, HubSpot, Stripe and Snowflake and custom MCP servers, keeping credentials out of the model's context.

Sourcedocs.digitalocean.com/products/inference/details/featuresread 2026-11-18

Workflow Orchestration Full

Agent routes send each query to the right agent based on its context, so a parent agent can sit in front of specialist agents for billing, product or technical questions, and function routes add calls to code within the same run. Workspaces group agents that work together and share evaluations. The Agent Development Kit, in public preview, deploys agent workflows written in Python frameworks as hosted endpoints with their steps traced.

Sourcedocs.digitalocean.com/products/inference/details/featuresread 2026-11-18

Knowledge Grounding & RAG Full

Knowledge bases are private retrieval stores with source data in Spaces and indexes in OpenSearch. They ingest Spaces buckets, local files, seed or sitemap URLs, Dropbox folders and Amazon S3 buckets, let teams pick the embedding model and a section based, semantic, hierarchical or fixed length chunking method, and retrieve by semantic, keyword or hybrid search with reranking. Auto indexing picks up changed sources, a RAG Playground tests retrieval, answers carry citations, and one agent can attach several knowledge bases.

Sourcedocs.digitalocean.com/products/inference/details/featuresread 2026-11-18

Human Oversight & Guardrails Partial

Guardrails scan every agent input and output for sensitive data, jailbreak attempts and harmful content and block what they catch, which filters content without asking a person. Approval of individual tool calls arrives with Managed Agents through the Cuttlefish iOS app, where a person follows a session and approves what the agent wants to do, but both are public previews that DigitalOcean advises against using for sensitive workloads. Managed agents on the generally available platform run without an approval step.

Sourcedocs.digitalocean.com/products/managed-agents/index.html.mdread 2026-11-18

Security, Identity & Governance Full

DigitalOcean holds SOC 2 and SOC 3 reports and states eligibility for HIPAA and DORA workloads, with certification and access reports through its trust portal. Every team member has a role, from six predefined roles to custom roles built from only the permissions chosen, and agents sit in projects and workspaces that decide who can see and share them. Agent endpoints require access keys unless switched to public, and serverless model access keys can be limited to a VPC network.

Sourcedigitalocean.com/trustread 2026-11-18

Observability & Auditability Full

Traceability lays out an agent's reasoning chain: the user input, each model output and the result of every function call, beside agent metrics and logs in the control panel. User and developer feedback ratings sit with each agent, and knowledge bases keep activity logs of indexing. With the Agent Development Kit in preview, traces are captured automatically and open from the CLI or the console in real time, and full conversation logs are still listed as coming soon.

Sourcedigitalocean.com/products/gradient/platformread 2026-11-18

Memory & State Persistence Partial

An agent carries conversational memory within a session as one of its stateful parts, alongside its data sources. In public preview, Managed Agents can pause, resume and fork sessions with their state intact, and persistent workspaces keep files in the workspace folder after a session ends. None of these is a memory layer with its own scope, retention and delete controls that an agent reads across separate conversations.

Sourcedigitalocean.com/products/gradient/platformread 2026-11-18

Deployment & Data Residency Full

A datacenter region is part of creating every agent, and the Agent Platform runs in Toronto (TOR1), Atlanta (ATL1) and Richmond (RIC1), so a team can keep agents in Canada or the US. Dedicated inference and bring your own models add New York (NYC2). Agents run next to the customer's Droplets, databases and Spaces buckets on the same DigitalOcean network, and nothing in the agent lineup runs outside North America today.

Sourcedocs.digitalocean.com/products/inference/details/availabilityread 2026-11-18

Prebuilt Agents / Templates / Packs Full

Agent templates start a build from a predefined use case such as customer support, through guided no code steps in the control panel, and instruction templates supply a starting role, goals, expertise and limits for the agent's instructions. DigitalOcean's tutorials add ready agent templates such as a SQL agent that answers questions over a database and a Telegram bot wired to an agent, which teams copy into their own account and adjust.

Sourcedocs.digitalocean.com/products/inference/details/featuresread 2026-11-18

Triggers & Channel Coverage Partial

Each agent is called through its own API endpoint with an access key, or made public without one, and through a chatbot snippet embedded in a website or app, with the playground for testing. Work reaches the agent when the customer's application or a website visitor sends a request. The managed Agent Platform has no schedule, webhook listener or event subscription that starts an agent on its own, so timed or event driven runs come from the customer's code calling the endpoint.

Sourcedocs.digitalocean.com/products/inference/details/featuresread 2026-11-18

Model Flexibility & Routing Full

Whoever builds the agent picks the provider and model, compares candidates in the Model Playground first, and can switch models after creation. The catalog spans OpenAI's GPT family, Anthropic's Claude family and open models such as Mistral, Llama and DeepSeek, and teams can bring their own OpenAI or Anthropic API key so the provider bills them directly. For direct inference calls, an Inference Router can pick the best model for each request instead.

Sourcedocs.digitalocean.com/products/gradient-ai-platform/how-to/create-agentsread 2026-11-18

APIs / SDKs / MCP Extensibility Full

The DigitalOcean API creates and manages agents through its gen AI agents endpoints, the CLI covers the same work, and each deployed agent has its own API endpoint with access keys. Serverless inference works with the OpenAI SDK, and a Knowledge Bases MCP server lets other agents query, filter and retrieve chunks. The Agent Development Kit, an Apache 2.0 Python SDK and CLI in public preview, deploys agent code and opens its logs and traces from the terminal.

Sourcedocs.digitalocean.com/products/gradient-ai-platform/how-to/create-agentsread 2026-11-18

Testing, Debugging & Optimization Full

Agent Evaluations work like a test suite for an agent: teams define prompts, pick metrics for response quality, safety, use of retrieved context, tool use and goal completion, and rerun them after changes to instructions, models, knowledge or routing on a schedule they set. Results come back as pass and fail percentages with charts, and a workspace can evaluate several agents at once. Evaluation runs bill at the model's token rates, and the agent playground covers hands on testing before release.

Sourcedocs.digitalocean.com/products/inference/details/featuresread 2026-11-18

Browser / Computer-use Partial

Browser control arrives with Managed Agents: its Harness Runtime sandboxes ship with Chromium for the agent to drive, and the Action Gateway includes DigitalOcean maintained browser automation tools alongside web search and web fetch. Both are in public preview, open to every customer but not recommended for production or sensitive workloads. The generally available Agent Platform reaches outside data through functions and knowledge bases rather than a browser.

Sourcedocs.digitalocean.com/products/managed-agents/index.html.mdread 2026-11-18

The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded

Pricing

Agent creation free; pay per token, such as GPT-5 mini at $0.25 per million input and $2 per million output tokens

Model input, output and cache tokens per million, plus knowledge base embedding and reranking tokens, guardrail tokens, Functions and OpenSearch storage

Included quota

Knowledge base indexing is charged only when files or URLs change, priced per million tokens and billed per thousand.

What is public

DigitalOcean charges nothing to create an agent and bills model tokens at serverless rates. GPT-5 mini costs $0.25 per million input tokens and $2 per million output tokens, Claude Haiku 4.5 costs $1 and $5, Claude Sonnet 4.6 costs $3 and $15, and GPT-5.5 costs $5 and $30.

Billing mechanics

Agent costs land on the same DigitalOcean invoice as other cloud resources, even when an agent switches models. Teams that bring their own OpenAI or Anthropic key pay the provider directly for those tokens.

Cost watchouts

Knowledge base storage bills through OpenSearch at its own rates, evaluations spend tokens at model prices, and functions an agent calls bill at Functions rates. Batch inference runs up to 50% cheaper for OpenAI and Anthropic models.

Variable cost rationale

Every part of an agent bills by usage, so cost follows conversation volume, model choice, retrieval and evaluation runs.

Additional watchouts

Managed Agents and the Agent Development Kit are in public preview, and DigitalOcean advises against running sensitive workloads on preview products.

Sales call required

No, self serve available

Free / trial

Agent creation and guardrail setup are free; usage bills from the first token

Lowest paid plan

No plans; agents bill model tokens, knowledge base embeddings, guardrails and functions as used

Commercial notes

Knowledge base embeddings cost $0.009 to $0.09 per million tokens depending on the model, reranking costs $0.01 per million tokens, and guardrails cost $0.20 per million tokens for content moderation and jailbreak detection and $0.34 for sensitive data detection.

Missing data

DigitalOcean gives no hosting price for agents deployed with the Agent Development Kit and no free credit amount for agents.

Agentic Index verified 2026-11-18

Alternatives to DigitalOcean

Other products that do the work DigitalOcean does, and what each one is.

  • Alteryx
  • HaystackOpen source Python framework by deepset for agents, RAG and search pipelines
  • PipecatOpen-source framework for building real-time voice and multimodal AI agents
  • ZenML
  • AgnoPython agent framework and AgentOS runtime
  • LiveKit

Contact us

Found a vendor we missed? Have feedback on the index? We'd love to hear from you.