Claude Self-Hosted

What Claude self-hosted sandboxes are

Anthropic shipped self-hosted sandboxes for Claude Managed Agents on 19 May 2026 (public beta). The split is clean: the agent loop — orchestration, context management, error recovery — stays on Anthropic’s infrastructure. Tool execution moves to a sandbox you point at. A worker process you run polls Anthropic’s work queue, claims sessions as they arrive, runs the tool calls, and posts results back.

The full docs are at platform.claude.com/docs/en/managed-agents/self-hosted-sandboxes. The beta lives behind the anthropic-beta: managed-agents-2026-04-01 header. Anthropic named four launch partners — Cloudflare, Daytona, Modal, and Vercel — and explicitly allows any infrastructure provider that can run the worker. We’re the fifth option, and the cheapest.

How it works on our side

• You create a self_hosted environment in your Anthropic Console and generate an ANTHROPIC_ENVIRONMENT_KEY.
• You give us the environment ID + key (or we walk you through the Console step on a kick-off call — your call).
• We provision a worker fleet on our Kathmandu GPUs. The ant beta:worker poll process (or the SDK’s EnvironmentWorker) sits on the work queue, claims sessions as Anthropic enqueues them, downloads skills into /workspace/skills/<name>/, runs tool calls, writes outputs to /mnt/session/outputs, posts results back.
• We scale the worker pool from work.stats queue depth — sessions never wait on idle capacity, idle workers never bill against your account.

Sample integration

# Your code, talking to Claude with our environment_id wired in:
client.beta.sessions.create(
    agent=my_agent.id,
    environment_id="env_…",     # the one we provisioned for you
    metadata={"input_file": "s3://your-bucket/data.csv"},
)

That’s the same shape Anthropic’s docs show for any self-hosted environment — your code doesn’t need to know we’re the host underneath.

Why us, alongside Cloudflare, Daytona, Modal, and Vercel

We undercut the four launch partners on price. Published comparison (numbers from each vendor’s pricing page or analyst write-ups, May 2026):

At EUR 0.12 we’re 20–50 % below the four named launch partners on a like-for-like sandbox shape. Idle workers aren’t billed (matches Vercel’s active-CPU model; better than Modal’s premium and Daytona’s wall-time billing).

Jurisdiction. Workloads run in Kathmandu, contracted through either Scalabs UG (Bavaria, German law, GDPR-aligned) or ScaLabs Cloud Pvt. Ltd. (Kathmandu, Nepali law). Nepali regulated entities can keep customer data on Nepali soil and still use Claude.

Sustainability. Nepal’s grid is predominantly hydropower; tool execution is meaningfully lower-carbon than the equivalent US/EU region. Measurable in your sustainability reporting.

No training on customer data. Same posture as the rest of ScaLabs Cloud — written into the contract, not just the FAQ.

Pre-installed sister APIs. Workers can call our LLM inference, Whisper STT, GLM-OCR / Mistral OCR document vision, and Chatterbox / Qwen3-TTS voice-cloning endpoints over the private Kathmandu network without leaving the perimeter — useful for agents that mix Claude reasoning with cheap utility-model calls.

Operations

• We expose work.stats queue-depth metrics, plus a passthrough so your own ops dashboards can read backlog and worker count alongside the rest of your monitoring.
• Stop-session is supported (graceful or force) via Anthropic’s work.stop endpoint.
• Skills (Anthropic’s agent skill-pack format) are downloaded fresh per session — no persistent skill cache leaks between tenants.
• Memory feature is not yet supported with self-hosted sandboxes (Anthropic restriction). When it ships we’ll add support.

Limits and what’s still beta

• Public beta on the Anthropic side; we are in early access. Pricing is founding-rate and subject to small adjustment before GA.
• Not available on Claude Platform on AWS (Anthropic restriction).
• HIPAA BAA / ZDR posture inherits from your Anthropic plan; we cover the sandbox-execution layer on our side under the same dual-entity contracting we use elsewhere.

Credit

Worker CLI and SDKs by Anthropic. We are the hosting provider; the agent loop and orchestration stay on Anthropic’s infrastructure.