Runtime
Sandbox
Execute tools and generated code through Docker, E2B, or Boxlite.
FastClaw runs agent tools in a sandbox when sandboxing is enabled. The sandbox receives a hydrated workspace, mounted skills, selected environment variables, and the configured tool policy.
Backends
| Backend | Use when | Notes |
|---|---|---|
docker | Local or self-hosted deployments | Uses long-lived or per-turn containers; supports bind mounts |
e2b | Cloud sandbox execution | Exposes sandbox ports through E2B URLs |
boxlite | Boxlite-compatible sandbox execution | Uses Boxlite API settings |
Bootstrap Env
FASTCLAW_SANDBOX_ENABLED
FASTCLAW_SANDBOX_BACKEND
FASTCLAW_SANDBOX_IMAGE
FASTCLAW_SANDBOX_BOXLITE_URL
FASTCLAW_SANDBOX_BOXLITE_CLIENT_ID
FASTCLAW_SANDBOX_BOXLITE_PREFIXDashboard runtime settings can also control sandbox enablement, backend, image/template, and network policy.
Workspace Hydration
Before a tool call, FastClaw hydrates:
- the chat/project workspace
- global skills under
$FASTCLAW_HOME/skills - agent-private skills under
$FASTCLAW_HOME/agents/<agentId>/agent/skills - selected environment variables and credentials
After execution, changed files are synced back to the durable store.
Policy
Tool execution is governed by policy. Keep production agents restrictive by default:
- limit network access when possible
- avoid passing broad secrets into the sandbox
- prefer per-agent or per-tool scoped credentials
- use object storage for distributed file hydration instead of local-only paths
Standalone Sandbox Commands
fastclaw sandbox create --image thinkany/fastclaw-sandbox:latest
fastclaw sandbox list
fastclaw sandbox connect <container-id>
fastclaw sandbox destroy <container-id>