A self-hosted alternative to Claude Managed Agents

Claude Managed Agents suits teams that want Anthropic to run the agent loop, the sandbox and the scaling, and that only need Claude. Agenhood suits teams that need agents, files and keys on their own server, or want to mix Claude Code with Codex, Opencode and their own tool loop.

Last updated

If you are looking for a Claude Managed Agents alternative, the first question is where the agent should live. Claude Managed Agents is Anthropic's hosted agent harness: you define an agent, and Anthropic runs the loop, the tools and the sandbox. Agenhood is an open-source platform you run on your own server with plain Docker, where each agent is a long-lived container. Agenhood also runs Claude Code as one of its drivers, so this is less a rivalry than two ways to use Anthropic's models.

The short answer

Pick Claude Managed Agents if

  • You don't want to run or patch any infrastructure.
  • You only use Claude models.
  • You want Anthropic to handle scaling, sandboxes and support.

Pick Agenhood if

  • Agents, files and keys must stay on your own server.
  • You want to mix Claude Code, Codex, Opencode and your own tool loop.
  • You want schedules, workflows and a console that also run on your server.
  • You want MIT-licensed code you can read and change.

Side by side

Topic Claude Managed Agents Agenhood
Who runs it Anthropic. You call its API with a Claude API key. source You, on one VM with Docker Compose. No Kubernetes.
Where the agent loop runs On Anthropic's side, with either environment type. source Inside each agent's container on your server.
Where tools run In an Anthropic-managed container per session, or on a worker you run (self-hosted sandbox). Web search and fetch run on Anthropic's side. source In the same hardened container. A filtering egress proxy is the only way out. The Vanilla driver searches the web through a self-hosted SearXNG.
Models Claude models, version 4.5 and later. source Anthropic and OpenAI models, through five drivers chosen per agent.
Persistent workspace Sessions keep their filesystem and history server-side and resume after pauses. source A /workspace volume per agent that survives restarts, pauses and weeks of idle. Without a linked Git repo, it is snapshotted after every task.
Scheduling Scheduled deployments: a cron expression and an IANA timezone. source One-time or recurring (hourly, daily, weekly, monthly), timezone-aware, with a calendar of upcoming runs.
Workflows Subagents, and dynamic workflows the agent writes itself to run many agents in phases. source Ordered chains of tasks you define across one or more agents. A schedule can fire a whole workflow.
Console Claude Console: build and test agents, plus a session viewer with timeline, tool calls and cost. source Web console with a fleet view, live task streams, usage dashboard and an in-browser terminal.
API style REST API with server-sent events. In beta behind a beta header. source REST API plus server-sent events. Named, revocable API keys.
Pricing model Model tokens plus a session runtime fee while a session is running. See their pricing page. source Free. You pay for your server and for model usage on your own keys.
License Hosted service that Anthropic operates. source MIT. The full source is on GitHub.

Where the agent loop, the tools and your data live

With Claude Managed Agents, Anthropic runs the agent loop. By default, tools run in a cloud environment: each session gets a fresh Linux container on Anthropic's infrastructure, with network rules you set. If tools must run on your machines, a self-hosted sandbox moves the bash and file tools to a worker you run, while Claude keeps running on Anthropic's side.

That split matters for data. Anthropic's guide to choosing an environment says that with either option the conversation history and tool results go to the Managed Agents service, and that web search and web fetch run on Anthropic's side. The overview adds that sessions store history, sandbox state and outputs server-side, which is why Managed Agents is not currently eligible for Zero Data Retention or a HIPAA BAA.

In Agenhood, the loop runs on your server. The driver runs inside the agent's container, tools run next to it, and task history, files and logs are stored in your own database and Docker volumes. The only way out is the egress proxy, which blocks private IP ranges and cloud metadata endpoints. Model calls to Anthropic or OpenAI still carry your prompts and tool output, so the difference is where things run and where they are stored, not what the model provider sees.

Which models you can use

A Managed Agents agent is configured with a Claude model, 4.5 or later. If Claude is all you use, that is a strength: Anthropic tunes the harness for its own models, with prompt caching and compaction built in.

Agenhood picks a driver per agent, behind one API: Vanilla (its own tool loop), Opencode, Codex, Claude Code (the unmodified claude -p) and API (a single LLM call per task, up to 32 in parallel). You can run a Claude Code agent next to a Codex agent and send both tasks the same way. Anthropic and OpenAI keys are stored encrypted (AES-GCM) on your server and never sent to the browser.

Schedules, workflows and the console

Both products cover these, in different shapes. Managed Agents has scheduled deployments: a cron expression and a timezone that start a new session on each run, with a run history. Its multiagent features let an agent hand work to subagents, or write a dynamic workflow that runs many agents in phases and combines the results. The Claude Console lets you build and test agents and has a session viewer with a timeline, tool calls and cost.

In Agenhood, schedules are one-time or recurring (hourly, daily, weekly or monthly) in your timezone, with a calendar of upcoming runs. Workflows are chains you define: ordered tasks across one or more agents, with a pipeline view. The console shows the whole fleet and streams every task live over server-sent events, with a token meter and a cancel button, and it opens a terminal into any agent. A GitHub issue or PR comment, or a Slack message, can also start a task, with the result edited back into the same thread.

What Claude Managed Agents does better

  • There is nothing to run. Anthropic operates the service, the sandboxes and their lifecycle. With Agenhood, the VM, the updates and the backups are yours.
  • The harness comes from the model's makers. Prompt caching, compaction and other tuning are built in, and you don't maintain an agent loop.
  • Agents can plan large jobs on the server. A dynamic workflow lets one agent start many others in phases and combine what they return. Agenhood workflows are steps you define ahead of time.
  • Hard spend caps. You can set a budget on a session or on every scheduled run, and the session pauses when it reaches the cap.

Moving over, or using both

The pieces map closely. A Managed Agents agent becomes an Agenhood agent with a driver and a model. A session becomes a task sent to that agent. A scheduled deployment becomes a scheduled task. MCP servers and skills from Git repos attach to Agenhood agents too.

Here is the weekly scan from Anthropic's scheduling docs, rebuilt on Agenhood with the Claude Code driver. With make dev running and an Anthropic API key added in the console, create the agent, save the prompt and schedule it. Use the id from the first two responses as $AGENT_ID and $PROMPT_ID.

Terminal
# Create a Claude Code agent
curl -X POST http://localhost:5173/v1/containers \
  -H "Authorization: Bearer tk_live_seedkey" \
  -H "Content-Type: application/json" \
  -d '{"name": "weekly-scan", "config": {"driver": "claude-code", "model": "sonnet"}}'

# Save the prompt it should run
curl -X POST http://localhost:5173/v1/prompts \
  -H "Authorization: Bearer tk_live_seedkey" \
  -H "Content-Type: application/json" \
  -d '{"name": "Weekly compliance scan", "body": "Run the weekly compliance scan and write the findings to report.md"}'

# Run it every Friday at 20:00, New York time
curl -X POST http://localhost:5173/v1/scheduled-tasks \
  -H "Authorization: Bearer tk_live_seedkey" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "Weekly compliance scan",
    "target": {"kind": "prompt", "container_id": "'"$AGENT_ID"'", "prompt_id": "'"$PROMPT_ID"'"},
    "schedule": {"kind": "recurring", "unit": "week", "time": "20:00", "weekdays": [5]},
    "timezone": "America/New_York"
  }'

Using both is just as reasonable. Keep agents that don't touch sensitive systems on Managed Agents, and run the ones that need your network, your files or OpenAI models on Agenhood. When they use Claude, both bill model usage to your own Anthropic account.

What it doesn't do

  • It is a hardened container, not a VM. Agents share the host's kernel. If you need a stronger boundary, give Agenhood its own VM and keep other workloads off it.
  • You run it. There is no hosted version, so updates, backups and uptime are on you.
  • It is built for a trusted team on your own server. Don't expose it as a public, multi-tenant service.

Questions people ask

Is Agenhood built on Claude Managed Agents?

No. Agenhood is a separate, open-source project and doesn't call the Managed Agents API. It runs agents in Docker containers on your server, and one of its drivers runs the unmodified Claude Code CLI. So both are ways to use Anthropic's models, not two versions of the same product.

Doesn't the self-hosted sandbox option already keep my data on my servers?

Partly. With a self-hosted sandbox, the agent's files and processes stay on a worker you run, but Claude still runs on Anthropic's side, and the conversation history, tool results, skills and memory stores go to Anthropic. With Agenhood, task history, files and logs are stored on your server. Each model call still sends its prompt and context to the provider you picked.

Can I use Claude models on Agenhood?

Yes. Pick the Claude Code driver, which runs the real claude CLI, or the Vanilla driver with an Anthropic API key. Keys are stored encrypted on your server and usage is billed to your own key. You can also sign in with your own Claude plan, on your own server, but we recommend an API key for scheduled or heavy work.

Can I use both?

Yes. They don't conflict. A common split is Claude Managed Agents for work you want Anthropic to host, and Agenhood for agents that must stay on your network or use OpenAI models. When they use Claude, both bill model usage to your own Anthropic key or plan.

What does each one cost?

Claude Managed Agents bills model tokens plus a session runtime fee while a session is running. Their pricing page has the current rates. Agenhood is free and MIT licensed: you pay for the server you run it on and for model usage on your own keys.

Which one is more secure?

It depends on who you want running the sandbox. With Claude Managed Agents, Anthropic runs and hardens it. With Agenhood, each agent gets a hardened Docker container with an egress proxy, but containers share the host kernel and you are responsible for patching and securing that host.

Run it locally in one command

You need Docker and Python 3.12 or newer. make dev builds the agent image, starts the stack, runs migrations, creates a login and prints the console URL.

Terminal
git clone https://github.com/appssemble/agenhood.git
cd agenhood
make dev

Sources

  1. Claude Managed Agents overview
  2. Choosing an agent environment
  3. Self-hosted sandboxes
  4. Cloud environment setup
  5. Define your agent
  6. Scheduled deployments
  7. Multiagent orchestration
  8. Build in Console
  9. Inspect sessions and track usage
  10. Claude API pricing (Managed Agents section)