Sandbox Configuration

Code sandbox settings for secure remote code execution. Supports Daytona, E2B, and CubeSandbox platforms.

General

Variable	Default	Description
ENABLE_SANDBOX	false	Enable sandbox execution.
SANDBOX_PLATFORM	daytona	Sandbox platform: daytona, e2b, or cubesandbox.
SANDBOX_GREP_TIMEOUT	30	Sandbox grep command timeout in seconds.

Daytona

Variable	Default	Sensitive	Description
DAYTONA_API_KEY	(empty)	Yes	Daytona API key.
DAYTONA_SERVER_URL	(empty)	No	Daytona server URL.
DAYTONA_TIMEOUT	180	No	Command timeout in seconds (3 minutes).
DAYTONA_IMAGE	(empty)	No	Sandbox image/snapshot ID to use.
DAYTONA_AUTO_STOP_INTERVAL	5	No	Auto-stop interval in minutes.
DAYTONA_AUTO_ARCHIVE_INTERVAL	5	No	Auto-archive interval in minutes.
DAYTONA_AUTO_DELETE_INTERVAL	1440	No	Auto-delete interval in minutes (24 hours).

E2B

Variable	Default	Sensitive	Description
E2B_API_KEY	(empty)	Yes	E2B API key.
E2B_TEMPLATE	base	No	Sandbox template name.
E2B_TIMEOUT	3600	No	Sandbox timeout in seconds (1 hour).
E2B_AUTO_PAUSE	true	No	Pause sandbox on timeout instead of killing (preserves state).
E2B_AUTO_RESUME	true	No	Auto-resume paused sandbox on next activity.

CubeSandbox

CubeSandbox is supported through the native CubeSandbox Python SDK. LambChat uses CubeSandbox metadata to keep a stable user-to-sandbox binding: each user should have one running sandbox that is shared across conversations and sessions.

Variable	Default	Sensitive	Description
CUBE_API_URL	http://127.0.0.1:3000	No	CubeSandbox API base URL. For the local dev environment this is commonly http://127.0.0.1:13000.
CUBE_TEMPLATE	(empty)	No	CubeSandbox template ID used when creating sandboxes.
CUBE_TIMEOUT	3600	No	Sandbox timeout in seconds.
CUBE_PROXY_NODE_IP	(empty)	No	Proxy node IP used by the SDK to reach sandbox data-plane services.
CUBE_PROXY_PORT_HTTP	80	No	HTTP proxy port used by the SDK.
CUBE_SANDBOX_DOMAIN	cube.app	No	Sandbox domain suffix used by CubeSandbox proxy routing.
CUBE_REQUEST_TIMEOUT	120	No	SDK request timeout in seconds. Lower values fail faster on stale data-plane connections; higher values tolerate slower local starts.
CUBE_AUTO_PAUSE	true	No	Ask CubeSandbox to pause on timeout instead of killing when supported by the runtime.
CUBE_AUTO_RESUME	true	No	Ask CubeSandbox to auto-resume paused sandboxes when supported by the runtime.

Lifecycle Behavior

When SANDBOX_PLATFORM=cubesandbox, LambChat follows this order for a user:

1. Reuse the in-process cached sandbox if it is still healthy.
2. Reconnect to the MongoDB user_sandbox_bindings record.
3. If the binding is missing or unhealthy, list CubeSandbox instances with matching metadata.user_id and reuse a healthy running sandbox.
4. Create a new sandbox only when no healthy existing sandbox can be used.
5. Clean duplicate same-user running sandboxes while keeping the selected sandbox.

CubeSandbox may occasionally report a sandbox as running in the control plane even though the data plane returns 504 Gateway Time-out. LambChat treats that as unhealthy when it cannot create the session work directory or run commands, then falls back to another existing sandbox or creates a new one.

Production Migration

LambChat stores sandbox bindings per user and per platform. New records use:

• sandboxes.e2b
• sandboxes.cubesandbox

Older production records may only have top-level fields such as sandbox_id and sandbox_state. When SANDBOX_PLATFORM=e2b, those legacy top-level records are treated as E2B bindings for backward compatibility. On the next successful reuse, LambChat writes the same sandbox into sandboxes.e2b, so the existing E2B sandbox continues to be used and is not recreated.

Switching a user or deployment from e2b to cubesandbox creates or reuses a CubeSandbox binding under sandboxes.cubesandbox without overwriting sandboxes.e2b. Switching back to e2b reads the E2B slot again.

For production deployments that should keep using E2B, leave:

SANDBOX_PLATFORM=e2b

Do not set SANDBOX_PLATFORM=cubesandbox unless you intentionally want those users to start using CubeSandbox. The two platforms have separate sandbox IDs and separate lifecycle APIs.

Performance Notes

• A cold CubeSandbox create depends on the template and local runtime restore time. In local testing this can take around 20-30 seconds.
• New conversations for the same user should normally reuse the same sandbox and avoid cold start.
• Duplicate sandbox cleanup runs outside the critical cache-hit path, so reuse can return before background cleanup finishes.
• CubeSandboxBackend uses CUBE_TIMEOUT for command execution and caches parent-directory creation for repeated file writes.
• The session manager caches successfully prepared session work directories in the current backend process. After a backend restart, mkdir -p is run again and remains safe.

Console and API Checks

For the local CubeSandbox dev environment used during integration:

# CubeSandbox Web UI
open http://127.0.0.1:12088

# List sandboxes from the Cube API
curl http://127.0.0.1:13000/sandboxes

Useful checks:

• The selected sandbox should have metadata.user_id equal to the LambChat user ID.
• A user should normally have only one healthy running sandbox.
• If a sandbox is listed as running but command execution returns 504, kill it from CubeSandbox or let LambChat replace it on the next sandbox initialization.

Examples

Daytona (Self-hosted)

ENABLE_SANDBOX=true
SANDBOX_PLATFORM=daytona
DAYTONA_API_KEY=your_daytona_api_key
DAYTONA_SERVER_URL=https://daytona.example.com
DAYTONA_TIMEOUT=180

E2B (Cloud)

ENABLE_SANDBOX=true
SANDBOX_PLATFORM=e2b
E2B_API_KEY=your_e2b_api_key
E2B_TEMPLATE=base
E2B_TIMEOUT=3600

CubeSandbox (Local Dev)

ENABLE_SANDBOX=true
SANDBOX_PLATFORM=cubesandbox
CUBE_API_URL=http://127.0.0.1:13000
CUBE_TEMPLATE=tpl-your-template-id
CUBE_PROXY_NODE_IP=127.0.0.1
CUBE_PROXY_PORT_HTTP=11080
CUBE_SANDBOX_DOMAIN=cube.app
CUBE_TIMEOUT=3600
CUBE_REQUEST_TIMEOUT=120
CUBE_AUTO_PAUSE=true
CUBE_AUTO_RESUME=true

INFO

The DAYTONA_AUTO_*_INTERVAL settings control sandbox lifecycle management to optimize resource usage. Sandboxes are automatically stopped, archived, and eventually deleted based on these intervals.

TIP

For CubeSandbox, tune CUBE_REQUEST_TIMEOUT based on your local runtime. Shorter values make stale sandboxes fail faster; longer values reduce false failures during slow local restores.