# Give an agent a throwaway machine per job with no credentials inside

> Give an agent a throwaway machine per job with no credentials inside: 2 companies fully do it, 18 with a limit, 24 carry the tag. Every listing in Code execution sandboxes with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/code-sandboxes.ephemeral-task-sandbox/
- Markdown: https://www.anchorterminal.com/niches/code-sandboxes.ephemeral-task-sandbox/index.md (~2,100 tokens)
- Slim: https://www.anchorterminal.com/niches/code-sandboxes.ephemeral-task-sandbox/index.min.md (~430 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/code-sandboxes.ephemeral-task-sandbox/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

Two or three companies' docs show a full fit.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 2 | 18 | 24 | 15 | 8 |

## What counts as done

A sandbox is created, the task runs and the sandbox is destroyed, with no credential readable from inside the guest (checked by printing the environment and the config files), and outbound calls to the needed hosts succeed through a broker or proxy.

Measured, for our own tests: those outbound calls succeed within 2 calls of setup.

How agents ask for it: run this task in a fresh sandbox; do this without exposing my API keys; spin up a clean box for each request; make sure the agent can't leak my tokens

## Does the job

The docs show it meets the test above, and nothing we read says otherwise.

- **Cloudflare Sandbox SDK (Cloudflare, Inc.)**: A Container starts with enableInternet false and reaches only hostnames an outbound handler in the Worker intercepts, where the credential is added; a sandbox per job is one Durable Object name and destroy() ends it.
- **E2B (FoundryLabs, Inc.)**: allowInternetAccess and allow and deny lists are set at creation, stored secrets are injected by the egress proxy so the guest holds only a reference, and Sandbox.create takes an Idempotency-Key. Doozy runs every job in a fresh credential-free E2B sandbox.

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **Agent 37 Cloud (Agent 37 Inc.)**: Network is allow-all by default and can be set to deny-all or an allow list live, and exec takes a `name` for safe resends. No credential proxy: secrets reach the guest through `env`, where commands can read them.
- **Amazon Bedrock AgentCore Code Interpreter (Amazon Web Services, Inc.)**: Sessions end and are wiped, and the network mode (Sandbox, Public or VPC) and IAM execution role are set per code interpreter, not per session. AWS services are reached by role without keys, but no proxy brokers third-party credentials.
- **Blaxel Sandboxes (Blaxel, Inc.)**: Domain allow and deny lists and proxy secret injection keep keys out of the guest, but both are public preview, set only at creation, and egress is open without them. Error bodies carry retryable and safe_to_retry_request flags.
- **Celesto (formerly SmolVM) (Celesto AI, Inc.)**: internet_settings mode off or restricted denies or limits egress, but the default is open, policy is fixed at creation, and there is no credential proxy, so keys the code needs go inside the guest.
- **Daytona (Daytona Platforms Inc.)**: networkBlockAll, networkAllowList and domainAllowList can be set at create or updated while running, but Tier 1 and 2 cannot override the organisation restriction, and no credential proxy was found. A user called credits near useless before Tier 3.
- **DeltaBox Lite (DeltaBox (delta-box))**: deny_all and allowlist network modes, secret-path denial and scrubbed child environments limit what a command can reach, but it runs on the developer's own machine with no credential proxy.
- **Deno Sandbox (Deno Land Inc.)**: `secrets` with a hosts list leaves only a placeholder in the guest, and allowNet limits egress, but allowNet is set at creation, the default is disputed between pages, and raw TCP to Postgres is impossible (a user built netfence instead).
- **Freestyle (Freestyle Cloud Inc.)**: Firewall rules are required and Outbound TLS injects secrets so git credentials stay at the edge, but there is no idempotency key, and an exec answering 409 VM_NON_RESPONSIVE may already have run.
- **InstaVM (Checksum Labs, Inc.)**: A session egress policy allows package managers, domains and CIDRs, but access is open until set and no credential proxy exists, so keys the code uses live in the guest.
- **Maritime (Maritime AI, Inc.)**: A `network` field and private networks exist and the create is idempotent, but the only broker is the model-call proxy; other service keys go into `env` inside the guest.
- **Microsoft Execution Containers (Microsoft Corporation)**: Egress resolves to deny when omitted and allow rules or a caller-managed runtimeConfig.networkProxy route traffic, but the proxy is yours to run and the workload shares the developer's machine.
- **Modal Sandboxes (Modal Labs, Inc.)**: block_network, CIDR and domain allowlists (beta) are set at create, a `name` makes a retried create raise AlreadyExistsError, and a sidecar credential proxy (beta) keeps secrets out. Dasher spins a fresh Modal sandbox per Slack request with a gateway for secrets.
- **Morph Cloud (Morph Labs, Inc.)**: No egress policy or firewall was found in the docs text, and no credential proxy; an Idempotency-Key appears on key-minting routes, not exec.
- **SPACE (Perplexity AI)**: Egress profiles and policies deny by default and creates and commands take an Idempotency-Key, but credential injection is off by default, needs AWS Secrets Manager and has no page of its own.
- **Runloop Devboxes (Runloop AI, Inc.)**: Network policies with allow_all false and Agent Gateways that hand the devbox only a bound gateway token fit the job, but egress is unrestricted by default and the service is winding down.
- **Sandbox0 (Lougao, LLC)**: trafficRules allow or block egress and can be updated at runtime, and credentialRules inject headers, placeholders, TLS client certificates or SSH keys at the egress boundary. A claim has no idempotency key, so a timed-out claim retried may run twice.
- **Sprites (Fly.io, Inc.)**: A DNS allowlist set from outside and Connectors that hold the token in a gateway fit the job, but egress is open until a policy is applied, and one user found Sprites cut off from Fly's private 6PN mesh.
- **Vercel Sandbox (Vercel Inc.)**: networkPolicy deny-all and allowlists update at runtime and credentials brokering injects headers so secrets never enter the sandbox, but the default is allow-all and issue 363 shows the SDK retrying a command POST with no idempotency key, running it twice.

- not this job: Offscreen (Mark Phelps (individual)): No network policy for the guest and `[prepare] run` executes on the host with the caller's environment; it does not aim at credential isolation.
- not this job: Together Code Sandbox (Together Computer, Inc.): No outbound network controls yet (the experimental policy covers inbound only) and no credential broker, so a key the code needs sits in the guest with open egress.

## What stops agents finishing

- egress is open by default and the allow list can only be set at creation (Blaxel, Sprites, E2B, Vercel, Deno)
- no idempotency key on command execution, so a retried request runs the command twice (Vercel issue 363, Sandbox0 claims)
- credential proxy only rewrites HTTP headers, so a database over raw TCP still needs the secret inside the guest

## Around it

Before: personal-assistants.slack-mention, agent-harnesses.plan-task. After: agent-memory.store-result, personal-assistants.reply-in-chat.

## Where this comes from

Weight 15 of 100 in Code execution sandboxes: 8 of 53 mentions. Credential isolation is the reason most often given on Hacker News for choosing a hosted sandbox at all (Doozy, Dasher, the Deno Sandbox thread, the Sprites Tokenizer experiment).

As data: https://www.anchorterminal.com/api/v1/use-cases/code-sandboxes.ephemeral-task-sandbox.json, or the MCP tool get_use_case with id code-sandboxes.ephemeral-task-sandbox. Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
