fabro/docs/execution/environments.mdx

---
title: "Environments"
description: "Sandbox providers for workflow execution"
---

When an agent runs a shell command, edits a file, or searches code, it does so inside a **sandbox**. The sandbox is the execution environment for all tool operations — it controls where commands run, which files are visible, and how much isolation exists between the agent and the host.

Arc supports four sandbox providers. Each one implements the same interface (file I/O, command execution, grep, glob), so workflows run identically regardless of which provider you choose. The difference is in where and how the tools execute.

| Provider | Runs on | Isolation | Use case |
|---|---|---|---|
| `local` | Host machine | None | Development, trusted workflows |
| `docker` | Docker container | Container-level | Reproducible environments, untrusted code |
| `daytona` | Cloud VM | Full machine | CI/CD, team-shared runs, SSH debugging |
| `exe` | Cloud VM ([exe.dev](https://exe.dev)) | Full machine | Fast ephemeral VMs, lightweight cloud sandboxing |

## Choosing a provider

Set the sandbox provider via CLI flag, [run config TOML](/execution/run-configuration), or server defaults:

```bash
# CLI flag
arc run workflow.dot --sandbox local
arc run workflow.dot --sandbox docker
arc run workflow.dot --sandbox daytona
arc run workflow.dot --sandbox exe
```

```toml title="run.toml"
# Run config TOML
[sandbox]
provider = "daytona"
```

The precedence order is: CLI flag > run config TOML > server defaults > built-in default (`local`).

## Local

The local sandbox runs all tool operations directly on the host machine. It's the default and the simplest option — no setup required beyond the Arc binary itself.

### How it works

- **Working directory** — Set to the current directory (or `directory` from the run config). Arc creates it if it doesn't exist.
- **Commands** — Executed via `/bin/bash -c` in the working directory.
- **File operations** — Read and write directly to the host filesystem. Relative paths resolve against the working directory.
- **Cleanup** — No-op. Local sandbox doesn't create or destroy anything on cleanup.

### Environment variable filtering

The local sandbox filters sensitive environment variables before passing them to commands. Variables ending in `_API_KEY`, `_SECRET`, `_TOKEN`, `_PASSWORD`, or `_CREDENTIAL` are stripped. A safelist of common variables (`PATH`, `HOME`, `USER`, `SHELL`, `LANG`, `TERM`, `TMPDIR`, `GOPATH`, `CARGO_HOME`, `NVM_DIR`) is always passed through.

<Note>
The local sandbox offers no isolation. Agents can read and modify any file on the host. Use `docker`, `daytona`, or `exe` when running untrusted workflows or when you need a reproducible environment.
</Note>

## Docker

The Docker sandbox runs all tool operations inside a Docker container. The host working directory is bind-mounted into the container, so file changes are visible on both sides.

### Prerequisites

- Docker Engine running on the host
- The configured image available locally (or `auto_pull` enabled)

### How it works

- **Container lifecycle** — On `initialize()`, Arc pulls the image (if needed), creates a container with `sleep infinity`, and starts it. On `cleanup()`, Arc stops and removes the container.
- **Working directory** — The host working directory is bind-mounted at `/workspace` inside the container. All relative paths resolve against this mount point.
- **Commands** — Executed via `docker exec` with `/bin/bash -c` inside the container. Timeout and cancellation are supported.
- **File writes** — Use the Docker API's tar upload to avoid shell escaping issues with special characters.
- **Platform detection** — The container's `uname -r` is cached at startup.

### Configuration

The Docker sandbox is configured through the `DockerSandboxConfig`:

| Setting | Default | Description |
|---|---|---|
| `image` | `arc-agent:latest` | Docker image to use |
| `container_mount_point` | `/workspace` | Mount point inside the container |
| `network_mode` | `bridge` | Docker network mode |
| `extra_mounts` | `[]` | Additional `host:container` bind mounts |
| `memory_limit` | unlimited | Memory limit in bytes |
| `cpu_quota` | unlimited | CPU quota (microseconds per 100ms period) |
| `auto_pull` | `true` | Pull the image if not found locally |
| `env_vars` | `[]` | Additional `KEY=VALUE` environment variables |

### Preserving the container

By default, the container is destroyed when the run finishes. To keep it alive for debugging:

```bash
arc run workflow.dot --sandbox docker --preserve-sandbox
```

Or in the run config:

```toml title="run.toml"
[sandbox]
provider = "docker"
preserve = true
```

When preserved, Arc prints the container ID so you can reconnect with `docker exec -it <id> bash`.

## Daytona

The Daytona sandbox runs all tool operations inside a cloud-hosted VM managed by [Daytona](https://daytona.io). It provides full machine-level isolation, automatic git cloning, and SSH access for debugging.

### Prerequisites

- A `DAYTONA_API_KEY` environment variable
- A GitHub App configured via `arc setup` (for private repository cloning)

### How it works

- **Sandbox lifecycle** — On `initialize()`, Arc creates a Daytona sandbox (from an image or a snapshot), clones the current git repository into it, and waits until it's ready. On `cleanup()`, the sandbox is deleted.
- **Working directory** — Fixed at `/home/daytona/workspace`. The current repository is cloned there automatically.
- **Git clone** — Arc detects the local `origin` remote URL and current branch, converts SSH URLs to HTTPS, and clones into the sandbox. For private repositories, Arc uses a GitHub App Installation Access Token scoped to `contents: read` on the specific repository. Public repositories are cloned without credentials. If no git repo is detected, the working directory is created empty.
- **Commands** — Executed via the Daytona process API. Commands are base64-encoded and piped through `sh` to support pipes, environment variables, and other shell features.
- **Ephemeral** — Sandboxes are created with `ephemeral: true` and a unique timestamped name (e.g. `arc-20260305-142301-a3f2`).

### Snapshots

Snapshots let you pre-build an environment image so each run starts with dependencies already installed. If the named snapshot doesn't exist and a `dockerfile` is provided, Arc creates it automatically and polls until it's ready (up to 10 minutes).

```toml title="run.toml"
[sandbox]
provider = "daytona"

[sandbox.daytona]
auto_stop_interval = 60

[sandbox.daytona.snapshot]
name = "rust-dev"
cpu = 4
memory = 8
disk = 20
dockerfile = "FROM rust:1.85-slim-bookworm\nRUN apt-get update && apt-get install -y git ripgrep"
```

| Field | Description |
|---|---|
| `name` | Snapshot identifier. Reused across runs if it already exists. |
| `cpu` | CPU cores for the snapshot VM. |
| `memory` | Memory in GB. |
| `disk` | Disk in GB. |
| `dockerfile` | Dockerfile content for building the snapshot. Required when creating a new snapshot. |

If the snapshot already exists and is in `Active` state, Arc uses it directly. If it's in `Building` or `Pending` state, Arc polls with exponential backoff until it's ready.

### Labels

Attach key-value labels to sandboxes for filtering and identification in the Daytona dashboard:

```toml title="run.toml"
[sandbox.daytona.labels]
project = "arc"
env = "ci"
team = "platform"
```

When using server defaults, labels are merged — run config labels override default labels on key collisions.

### SSH access

Connect to a running Daytona sandbox via SSH for live debugging:

```bash
arc run workflow.dot --sandbox daytona --ssh
```

This creates temporary SSH credentials (valid for 60 minutes) and prints the connection command.

### Preserving the sandbox

Like Docker, Daytona sandboxes are destroyed on cleanup by default. Use `--preserve-sandbox` to keep them alive:

```bash
arc run workflow.dot --sandbox daytona --preserve-sandbox
```

Arc prints the sandbox name so you can find it in the [Daytona dashboard](https://app.daytona.io/dashboard/sandboxes).

### Auto-stop

The `auto_stop_interval` setting (in minutes) tells Daytona to stop the sandbox after a period of inactivity. This saves costs for long-running sandboxes that may sit idle:

```toml title="run.toml"
[sandbox.daytona]
auto_stop_interval = 30
```

## Exe

The Exe sandbox runs all tool operations inside a cloud VM managed by [exe.dev](https://exe.dev). It provides full machine-level isolation with fast VM startup via SSH.

### Prerequisites

- SSH keys configured for `exe.dev` (added via the exe.dev dashboard)
- A GitHub App configured via `arc setup` (for private repository cloning)

### How it works

- **VM lifecycle** — On `initialize()`, Arc connects to the exe.dev management plane via SSH and runs `new --json` to create a VM. The response contains the VM name and an SSH destination (e.g. `my-vm.exe.xyz`). On `cleanup()`, Arc runs `rm <vm_name>` to destroy the VM. Cancelling a run cleanly tears down the VM.
- **Working directory** — Fixed at `/home/exedev`. Relative paths resolve against this directory.
- **Git clone** — Arc detects the local `origin` remote URL and current branch, converts SSH URLs to HTTPS, and clones into the VM. For private repositories, Arc uses a GitHub App Installation Access Token. Public repositories are cloned without credentials. If no git repo is detected, the working directory is created empty.
- **Commands** — Executed via SSH on the data plane (`<vmname>.exe.xyz`). Commands are base64-encoded and piped through `sh` to handle pipes, environment variables, and other shell features. Environment variables and working directory overrides are supported.
- **File I/O** — Reads use `cat` over SSH. Writes use SCP upload. Parent directories are created automatically.
- **Git checkpointing** — Between stages, Arc commits agent changes and pushes them to a remote branch, the same mechanism used by Daytona sandboxes.
- **Ephemeral** — Each run gets a fresh VM that is destroyed on cleanup.

### Configuration

```toml title="run.toml"
[sandbox]
provider = "exe"

[sandbox.exe]
image = "my-custom-image:latest"
```

| Field | Description |
|---|---|
| `image` | Custom container image for the VM. Optional — uses the exe.dev default when omitted. |

### SSH access

Connect to a running exe.dev sandbox via SSH for live debugging:

```bash
arc run workflow.dot --sandbox exe --ssh
```

This prints the SSH connection command so you can connect to the VM while the workflow runs.

## Sandboxing

Sandboxes isolate agent execution from the host machine. When an agent runs a shell command, edits a file, or searches code, it does so inside a sandbox — preventing unintended side effects and providing a reproducible environment for each run.

### Filesystem

Each provider offers a different level of filesystem isolation:

| Provider | Isolation | What agents can access |
|---|---|---|
| `local` | None | The entire host filesystem. Agents can read and modify any file. |
| `docker` | Container-level | Only the bind-mounted working directory (`/workspace` by default) and whatever is in the container image. Host files outside the mount are inaccessible. |
| `daytona` | Full machine | A cloud VM with the repository cloned into `/home/daytona/workspace`. The host filesystem is completely inaccessible. |
| `exe` | Full machine | A cloud VM with the repository cloned into `/home/exedev`. The host filesystem is completely inaccessible. |

For `local`, Arc filters sensitive environment variables (those ending in `_API_KEY`, `_SECRET`, `_TOKEN`, `_PASSWORD`, or `_CREDENTIAL`) but does not restrict file access. Use `docker`, `daytona`, or `exe` when running untrusted workflows.

### Network

Each provider handles outbound network access differently:

| Provider | Default | Controls |
|---|---|---|
| `local` | Full access | No network isolation. Agents have the same network access as the host. |
| `docker` | Bridge network | Set via the `network_mode` config option. Supports all Docker network modes (`bridge`, `none`, `host`, etc.). |
| `daytona` | Full access | Configurable via the `network` setting with three modes: `"allow_all"`, `"block"`, or CIDR-based allow lists. |
| `exe` | Full access | No network isolation controls. VMs have full outbound access. |

For Daytona, network access is configured in the `[sandbox.daytona]` section:

```toml title="run.toml"
# Block all egress
[sandbox.daytona]
network = "block"

# Allow only specific CIDRs
[sandbox.daytona]
network = { allow_list = ["208.80.154.232/32", "10.0.0.0/8"] }
```

When using server defaults, the run config `network` overrides the server default. If neither specifies `network`, Daytona's own default (full access) applies.

## Safety guardrails

Regardless of which provider you use, Arc applies a **read-before-write** guardrail. Agents must read a file (via `read_file` or `grep`) before they can modify it with `write_file` or `delete_file`. Writing to new files that don't yet exist is always allowed. This prevents agents from blindly overwriting files they haven't inspected.