Set up
Sandboxes
A sandbox is an isolated computer where an agent runs commands and works with files, so it can do work like building charts, spreadsheets, and scripts.
Each conversation gets its own sandbox the first time the agent needs one. Files under /workspace persist across turns in that conversation for the sandbox's one-hour lifetime. A sandbox is a scratch space, not a file store: ask the agent to attach anything you want to keep to its reply.
Choose a provider
| Provider | Runs on | Connection needs |
|---|---|---|
| Adapt Platform | Adapt Platform sandboxes | Browser sign-in |
| Modal | Modal | Token ID and token secret |
| E2B | E2B | API key |
| Daytona | Daytona | API key |
| Local Docker | The Adapt server's Docker daemon | Nothing; self-hosted installs only |
Owners and admins add sandbox connections in Connections, under Sandboxes. Setup also offers one as an optional step.
Give an agent a sandbox
Choose the sandbox for new agents in Settings → Defaults → Sandbox, and for one agent in its Settings tab under Sandbox. Image and resource settings stay editable after you pick a connection. Anyone who can edit an agent can choose any enabled sandbox connection.
Settings → Defaults → Sandbox also controls availability for the whole workspace:
- Off for everyone: no agent can use a sandbox.
- Off by default: off unless turned on for an agent.
- On by default: copied into new agents. Existing agents keep their settings.
What agents can do in a sandbox
- Run command: run a shell command in
/workspace. Each command has a 30-second time limit by default, up to 10 minutes, and the run's remaining time can stop it sooner. - Read file and Write file: read UTF-8 files up to 64 KB, and write text files up to 32,000 characters, under
/workspace. - Copy attachment: copy a file from the conversation into
/workspace/attachments, to open spreadsheets, documents, and archives. - Attach file: attach a file the agent built to its reply, for you to download.
- Preview: show a web server running in the sandbox in a panel beside the chat. See Previews.
Model and integration credentials are never placed in a sandbox. To give sandbox commands an API key or token, use Secrets.
Sandbox lifetime
A sandbox expires one hour after it's created. The agent opens a new one when a later turn needs it, so files from an expired sandbox are gone.
Changing the agent's sandbox connection, its settings, or its selected secrets also opens a new sandbox on the next command. The old one keeps running until its original expiry.
If Adapt can't tell whether the provider created a sandbox, for example because the request timed out, it doesn't silently create another one for that conversation until the first would have expired. Check the provider's own dashboard when that happens.
Adapt Platform
Choose Sign in to Adapt Platform, sign in, and approve the Platform organization to use. Adapt stores the credentials encrypted and refreshes them automatically. API keys aren't supported.
When you sign in, choose a workspace or personal connection. Only a workspace connection can be an agent's sandbox. Selecting Adapt Platform for an agent gives it both a sandbox and the connected services approved for that connection in Platform; removing it removes both.
- Container image: any image Platform can pull. Unset uses Platform's default, currently
ubuntu:24.04. - CPU cores, Memory (MiB), Disk (MiB): unset uses Platform's defaults, currently 1 CPU, 1,024 MiB, and 2,048 MiB.
Platform sandboxes have open internet access. To revoke access, use Platform's Settings → Agent connections; choose Sign in again on the Adapt connection to restore it. Access stays tied to the person who approved it and their membership in the Platform organization.
Adapt connects to https://platform.adapt.com. A self-hosted install can point at another Platform deployment:
PLATFORM_URL=https://platform.adapt.com
Modal
- Token ID and Token secret: from your Modal account.
- Modal app name: defaults to
adapt. - Container image: defaults to
python:3.13-slim. - CPU request, Memory request, CPU limit, Memory limit, and GPU: optional resources.
- Block network access:
trueorfalse. Unset leaves Modal's default, which allows outbound internet access.
E2B
- E2B API key: from your E2B account.
- Sandbox template: unset uses E2B's
basetemplate. - Block network access:
trueorfalse. It limits outbound traffic only.
Daytona
- Daytona API key: from your Daytona account.
- Container image: defaults to
python:3.13-slim. - CPU, Memory, GPU count, and GPU type: unset uses the image's defaults and no GPU.
- Block network access:
trueorfalse. Unset leaves your Daytona organization's network policy in charge.
Local Docker
Local Docker runs sandboxes as containers on the Adapt server itself. It's off unless the server sets:
ENABLE_DOCKER_SANDBOX=true
The server needs the Docker CLI and permission to reach its Docker daemon. Use a machine dedicated to sandboxing. Pull the image before agents use it:
docker pull python:3.13-slim
- Preloaded container image: defaults to
python:3.13-slim. The image must includesh,sleep,timeout,head,mkdir, andcat. - Internet access: off by default. Turn it on to let agents install packages with
pip installornpm install. Commands can then reach the internet, the host network, and other containers on Docker's default bridge where Docker allows it.
Each container runs with:
- No network unless Internet access is on.
- A read-only root filesystem, no host mounts, dropped capabilities, and an unprivileged user.
- 1 CPU, 1 GiB of memory, and at most 128 processes.
- 256 MB for
/workspaceand 64 MB for/tmp, both in memory.
Use secrets in a sandbox
Secrets are environment values saved in Settings → Secrets. In an agent's sandbox settings, select the secrets its commands should get; Agent defaults can't select secrets. Selected values are set as environment variables when the sandbox is created and when each command runs. The agent is told their names and instructed never to read the values into the conversation.
Adapt removes known secret values from command output, file reads, and provider errors. That stops accidental leaks, not code written to send a secret somewhere. Only select a secret for an agent whose tools, sandbox provider, and image you trust with it. See Secrets.
Previews
An agent can start a web server in its sandbox and show it in a panel beside the chat. Who can open it depends on the provider:
- Adapt Platform: any port from 1024 to 65535, through a public Platform endpoint. Anyone with the URL can reach it while the sandbox runs.
- Modal: ports 3000, 5173, 8000, and 8080 only, through encrypted tunnels. Anyone with the URL can reach it. Blocking network access also blocks previews.
- E2B: any port from 1024 to 65535, at a public
e2b.appURL. Anyone with the URL can reach it. - Daytona: any port from 1024 to 65535, as a signed URL valid until the sandbox expires. Anyone with the link can reach it.
- Local Docker: any port from 1024 to 65535, published only on the server's
127.0.0.1, so only a browser on the Adapt server's machine can open it. Needs Docker Engine 28 or later.
Servers must listen on 0.0.0.0, not localhost. Local Docker previews also need setsid and tail in the image.