Self-host Claude Managed Agents

Claude Managed Agents runs Claude’s agent loop on Anthropic’s platform. With a self-hosted sandbox, the agent’s tools run on infrastructure you choose instead of Anthropic’s cloud sandbox. An Agent37 instance makes a good one: an always-on Linux computer that runs Anthropic’s environment worker as its main process, keeps the agent’s files on its own disk, and bills by the minute, from $4.76 a month. This is not the same as Host Claude Code. There, the whole agent runs on the instance and you talk to it through the Agent API. Here, Anthropic runs the agent, you talk to it through the Claude API, and the instance runs its tool calls.

Paste this into your coding agent

Read https://www.agent37.com/docs/llms-full.txt and https://platform.claude.com/docs/en/managed-agents/self-hosted-sandboxes.
I want my Claude Managed Agents sessions to run their tools on an Agent37 instance instead of Anthropic's cloud sandbox.
Build the worker image from the Dockerfile on the Agent37 "Self-host Claude Managed Agents" page with npx agent37 templates build . --name claude-managed-agents-worker, then create an instance from that template with ANTHROPIC_ENVIRONMENT_KEY and ANTHROPIC_ENVIRONMENT_ID in env.
Done when GET https://api.anthropic.com/v1/environments/{id}/work/stats shows workers_polling of 1 and a session on that environment runs a bash command on the instance.
My keys are in AGENT37_API_KEY, ANTHROPIC_API_KEY, ANTHROPIC_ENVIRONMENT_KEY, and ANTHROPIC_ENVIRONMENT_ID.

How it works

Anthropic runs the agent loop and the model. Each session you create on a self-hosted environment waits in that environment’s work queue.
Your Agent37 instance runs ant beta:worker poll, the environment worker from Anthropic’s ant CLI. It claims a session from the queue, runs the agent’s bash and file tools in /workspace, and posts the results back. It works one session at a time.
The worker only makes outbound HTTPS calls to api.anthropic.com. The instance needs no public URL, open port, or webhook.

Before you begin

An Agent37 API key, exported as AGENT37_API_KEY. Create one in the dashboard.
A Claude API key, exported as ANTHROPIC_API_KEY.
Node.js, for the npx agent37 CLI that builds the image. No local Docker is needed.

1. Create a self-hosted environment

Create the environment with the Claude API, or in the Claude Console under Environments > New > Self-hosted:

curl

curl https://api.anthropic.com/v1/environments \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "anthropic-beta: managed-agents-2026-04-01" \
  -H "content-type: application/json" \
  -d '{ "name": "agent37", "config": { "type": "self_hosted" } }'

Then open the environment in the Console and click Generate environment key. Key generation is Console-only. Export both values:

export ANTHROPIC_ENVIRONMENT_ID="env_..."
export ANTHROPIC_ENVIRONMENT_KEY="sk-ant-oat01-..."

The environment key is what the worker authenticates with. A Claude API key in its place fails with 401 Invalid bearer token.

2. Build the worker image

Put this Dockerfile in an empty folder:

Dockerfile

FROM node:22-bookworm-slim

# Tools the agent can reach for from its bash tool; add your own here.
RUN apt-get update && apt-get install -y --no-install-recommends \
      ca-certificates curl git python3 ripgrep tini \
 && rm -rf /var/lib/apt/lists/*

# The ant CLI ships Anthropic's environment worker.
ARG ANT_VERSION=1.35.0
RUN curl -fsSL "https://github.com/anthropics/anthropic-cli/releases/download/v${ANT_VERSION}/ant_${ANT_VERSION}_linux_amd64.tar.gz" \
    | tar -xz -C /usr/local/bin ant

# The agent's bash tool runs on a terminal: without this, `git log` waits on a pager until the tool times out.
ENV PAGER=cat

WORKDIR /workspace
ENTRYPOINT ["tini", "--", "ant", "beta:worker", "poll", "--workdir", "/workspace"]

The worker is the image’s main process, and every bash command the agent issues runs inside this image, so install whatever your agent needs: a language runtime, a CLI, your internal packages. tini reaps the background processes those commands leave behind. The image runs on linux/amd64, which is why it fetches the amd64 build of ant; newer versions are on the ant releases page. Build it on Agent37 and publish it as a workspace template:

npx agent37 templates build . --name claude-managed-agents-worker

The build runs on Agent37’s infrastructure and streams its log to your terminal. See Build a custom image for build secrets, private base images, and updates.

3. Start the worker

Create an instance from the template, with the environment id and key as instance env:

curl

curl -X POST https://api.agent37.com/v1/instances \
  -H "Authorization: Bearer $AGENT37_API_KEY" \
  -H "Content-Type: application/json" \
  -d "{
    \"template\": \"claude-managed-agents-worker\",
    \"env\": {
      \"ANTHROPIC_ENVIRONMENT_ID\": \"$ANTHROPIC_ENVIRONMENT_ID\",
      \"ANTHROPIC_ENVIRONMENT_KEY\": \"$ANTHROPIC_ENVIRONMENT_KEY\"
    }
  }"

The instance is running within seconds, and the worker starts polling. Leave auto-sleep off: the worker’s outbound polling does not count as activity, so a sleeping instance stops claiming sessions. Confirm the worker is connected. workers_polling should read 1:

curl

curl https://api.anthropic.com/v1/environments/$ANTHROPIC_ENVIRONMENT_ID/work/stats \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "anthropic-beta: managed-agents-2026-04-01"

response

{ "type": "work_queue_stats", "depth": 0, "pending": 0, "oldest_queued_at": null, "workers_polling": 1 }

The worker’s own output is in the instance logs: idle; polling for work while it waits, then claimed work and executing tool lines once a session runs.

4. Run a session

Any Managed Agents agent works; nothing in the agent says where its tools run. If you don’t have one yet, create one with the built-in toolset:

curl

curl https://api.anthropic.com/v1/agents \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "anthropic-beta: managed-agents-2026-04-01" \
  -H "content-type: application/json" \
  -d '{
    "name": "Agent37 assistant",
    "model": "claude-sonnet-5",
    "system": "You are a helpful assistant.",
    "tools": [{ "type": "agent_toolset_20260401" }]
  }'

Start a session on your self-hosted environment, then send it a message:

curl

curl https://api.anthropic.com/v1/sessions \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "anthropic-beta: managed-agents-2026-04-01" \
  -H "content-type: application/json" \
  -d "{ \"agent\": \"$AGENT_ID\", \"environment_id\": \"$ANTHROPIC_ENVIRONMENT_ID\" }"

curl https://api.anthropic.com/v1/sessions/$SESSION_ID/events \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "anthropic-beta: managed-agents-2026-04-01" \
  -H "content-type: application/json" \
  -d '{ "events": [{ "type": "user.message", "content": [{ "type": "text", "text": "Run hostname and save the output to /workspace/hello.txt." }] }] }'

Read the turn back with GET /v1/sessions/$SESSION_ID/events, or stream it; see Anthropic’s events and streaming. The bash tool result shows the command ran on your instance, with the instance id as the hostname:

response

{
  "type": "user.tool_result",
  "content": [{ "type": "text", "text": "ab12cd34ef\n" }]
}

Work with the agent’s files

Everything the agent writes stays on the instance’s disk, across sessions and restarts. Read it from your backend with exec:

curl

curl -X POST https://sandbox.agent37.app/v1/instances/ab12cd34ef/exec \
  -H "X-Agent37-Key: $AGENT37_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "command": "cat /workspace/hello.txt" }'

Anthropic doesn’t mount files or GitHub repositories into a self-hosted sandbox, so stage inputs yourself: git clone or curl them into /workspace over exec before you start the session. To watch the agent work, SSH into the instance.

Scale out

More sessions at once. A worker holds one session until a minute after its turn ends (the --max-idle flag, default 1m), so the next message in a conversation continues without going back to the queue. Other sessions wait their turn. To run several at once, create more instances from the same template with the same env: every worker on an environment polls one queue, and each session goes to the first free worker. A later message on a session that was let go queues again and runs on whichever worker is free.
Heavier tools. Resize the instance for more CPU, memory, or disk.
Separate customers. Sessions on one instance share its disk. When customers must not see each other’s files, give each one its own self-hosted environment and instance.
Memory stores. The ant worker doesn’t mount memory stores. If your sessions attach them, replace the ENTRYPOINT with the Python, TypeScript, or Go SDK’s EnvironmentWorker; see Use memory stores.

One instance per session

To give every session its own computer, claim the work yourself instead of running ant beta:worker poll. A small orchestrator on your server long-polls the environment’s queue with the Anthropic SDK’s WorkPoller, creates an instance with auto_sleep: true for each new session, and starts ant beta:worker run on it over exec, which wakes the instance if it is asleep. The session’s files stay on its instance between turns, the instance sleeps and bills disk alone while nobody is talking to it, and sessions run in parallel, one per instance. The claude-managed-agents example is a complete orchestrator in about 100 lines of Node.js, with a chat script and a parallel-sessions demo.

Cost

The default 2 vCPU / 4 GB instance costs $4.76 a month running around the clock, metered per minute from your Agent37 wallet; larger shapes are on Instances. Model usage bills to your Claude account, not to Agent37.

Troubleshooting

Symptom	Fix
`workers_polling` stays `0`	Read the instance logs. `401 Unauthorized` with `Invalid bearer token` means `ANTHROPIC_ENVIRONMENT_KEY` is wrong or is a Claude API key. Env is set at create, so delete the instance and create it again with the right key.
`restart_count` climbs in the logs health readout	The worker exits on a permanent error and the platform starts it again. The error is at the top of the logs.
A session stays queued	Every worker is busy, or none is polling. Check that `workers_polling` is at least `1`, and add instances if sessions pile up in `depth`.
A turn stalls after the instance restarts mid-session	The stopped worker still held the session. Anthropic releases it within a few minutes, and the session then runs on the next free worker. Restart, resize, or update between sessions to avoid the wait.

Get started

Sandboxes

Agent templates

Build with Agent37

Self-host Claude Managed Agents

How it works

Before you begin

1. Create a self-hosted environment

2. Build the worker image

3. Start the worker

4. Run a session

Work with the agent’s files

Scale out

One instance per session

Cost

Troubleshooting

​How it works

​Before you begin

​1. Create a self-hosted environment

​2. Build the worker image

​3. Start the worker

​4. Run a session

​Work with the agent’s files

​Scale out

​One instance per session

​Cost

​Troubleshooting

How it works

Before you begin

1. Create a self-hosted environment

2. Build the worker image

3. Start the worker

4. Run a session

Work with the agent’s files

Scale out

One instance per session

Cost

Troubleshooting