Merge pull request #1904 from dgageot/local-sandbox

cagent run --sandbox

Author: dgageot

Dockerfile

@@ -50,6 +50,10 @@ COPY --from=builder /binaries/cagent-$TARGETOS-$TARGETARCH* cagent
 FROM scratch AS cross
 COPY --from=builder /binaries .
 
+FROM docker/sandbox-templates:cagent AS sandbox-template
+ARG TARGETOS TARGETARCH
+COPY --from=builder /binaries/cagent-$TARGETOS-$TARGETARCH /usr/local/bin/cagent
+
 FROM alpine:${ALPINE_VERSION}
 RUN apk add --no-cache ca-certificates docker-cli && \
     addgroup -S cagent && adduser -S -G cagent cagent && \

Taskfile.yml

@@ -88,6 +88,10 @@ tasks:
     desc: Build and Push Docker image
     cmd: docker buildx build --push --platform linux/amd64,linux/arm64 -t docker/cagent -t docker/docker-agent  {{.BUILD_ARGS}} .
 
+  build-sandbox-template:
+    desc: Build a sandbox template with the local cagent binary
+    cmd: docker buildx build --target=sandbox-template -t cagent-dev {{.BUILD_ARGS}} --load .
+
   record-demo:
     desc: Record demo gif
     cmd: vhs ./docs/recordings/demo.tape

agent-schema.json

@@ -831,10 +831,6 @@
           "type": "string",
           "description": "Name for the a2a tool"
         },
-        "sandbox": {
-          "$ref": "#/definitions/SandboxConfig",
-          "description": "Sandbox configuration for running shell commands in a Docker container (shell tool only)"
-        },
         "file_types": {
           "type": "array",
           "description": "File extensions this LSP server handles (e.g., [\".go\", \".mod\"]). Only for lsp toolsets.",
@@ -1012,48 +1008,6 @@
       ],
       "additionalProperties": false
     },
-    "SandboxConfig": {
-      "type": "object",
-      "description": "Configuration for running shell commands inside a sandboxed Docker container",
-      "properties": {
-        "image": {
-          "type": "string",
-          "description": "Docker image to use for the sandbox container. Defaults to 'alpine:latest' if not specified.",
-          "examples": [
-            "alpine:latest",
-            "ubuntu:22.04",
-            "python:3.12-alpine",
-            "node:20-alpine"
-          ]
-        },
-        "paths": {
-          "type": "array",
-          "description": "List of paths to bind-mount into the container. Each path can have an optional ':ro' suffix for read-only access (default is read-write ':rw'). Relative paths are resolved from the agent's working directory.",
-          "items": {
-            "type": "string"
-          },
-          "minItems": 1,
-          "examples": [
-            [
-              ".",
-              "/tmp"
-            ],
-            [
-              "./src",
-              "./config:ro"
-            ],
-            [
-              "/data:rw",
-              "/secrets:ro"
-            ]
-          ]
-        }
-      },
-      "required": [
-        "paths"
-      ],
-      "additionalProperties": false
-    },
     "ScriptShellToolConfig": {
       "type": "object",
       "description": "Configuration for custom shell tool",

cmd/root/run.go

@@ -49,6 +49,8 @@ type runExecFlags struct {
 	cpuProfile        string
 	memProfile        string
 	forceTUI          bool
+	sandbox           bool
+	sandboxTemplate   string
 
 	// Exec only
 	exec          bool
@@ -111,6 +113,8 @@ func addRunOrExecFlags(cmd *cobra.Command, flags *runExecFlags) {
 	_ = cmd.PersistentFlags().MarkHidden("memprofile")
 	cmd.PersistentFlags().BoolVar(&flags.forceTUI, "force-tui", false, "Force TUI mode even when not in a terminal")
 	_ = cmd.PersistentFlags().MarkHidden("force-tui")
+	cmd.PersistentFlags().BoolVar(&flags.sandbox, "sandbox", false, "Run the agent inside a Docker sandbox (requires Docker Desktop with sandbox support)")
+	cmd.PersistentFlags().StringVar(&flags.sandboxTemplate, "template", "", "Template image for the sandbox (passed to docker sandbox create -t)")
 	cmd.MarkFlagsMutuallyExclusive("fake", "record")
 
 	// --exec only
@@ -120,6 +124,11 @@ func addRunOrExecFlags(cmd *cobra.Command, flags *runExecFlags) {
 }
 
 func (f *runExecFlags) runRunCommand(cmd *cobra.Command, args []string) error {
+	// If --sandbox is set, delegate everything to `docker sandbox run cagent`.
+	if f.sandbox {
+		return runInSandbox(cmd, &f.runConfig, f.sandboxTemplate)
+	}
+
 	if f.exec {
 		telemetry.TrackCommand("exec", args)
 	} else {

cmd/root/sandbox.go

@@ -0,0 +1,76 @@
+package root
+
+import (
+	"cmp"
+	"errors"
+	"fmt"
+	"log/slog"
+	"os"
+	"os/exec"
+
+	"github.com/spf13/cobra"
+
+	"github.com/docker/cagent/pkg/config"
+	"github.com/docker/cagent/pkg/environment"
+	"github.com/docker/cagent/pkg/paths"
+	"github.com/docker/cagent/pkg/sandbox"
+)
+
+// runInSandbox delegates the current command to a Docker sandbox.
+// It ensures a sandbox exists (creating or recreating as needed), then
+// executes cagent inside it via `docker sandbox exec`.
+func runInSandbox(cmd *cobra.Command, runConfig *config.RuntimeConfig, template string) error {
+	if environment.InSandbox() {
+		return fmt.Errorf("already running inside a Docker sandbox (VM %s)", os.Getenv("SANDBOX_VM_ID"))
+	}
+
+	ctx := cmd.Context()
+	if err := sandbox.CheckAvailable(ctx); err != nil {
+		return err
+	}
+
+	cagentArgs := sandbox.BuildCagentArgs(os.Args)
+	agentRef := sandbox.AgentRefFromArgs(cagentArgs)
+	configDir := paths.GetConfigDir()
+
+	// Always forward config directory paths so the sandbox-side
+	// cagent resolves it to the same host directories
+	// (which is mounted read-write by ensureSandbox).
+	cagentArgs = sandbox.AppendFlagIfMissing(cagentArgs, "--config-dir", configDir)
+
+	stopTokenWriter := sandbox.StartTokenWriterIfNeeded(ctx, configDir, runConfig.ModelsGateway)
+	defer stopTokenWriter()
+
+	// Ensure a sandbox with the right workspace mounts exists.
+	wd := cmp.Or(runConfig.WorkingDir, ".")
+	name, err := sandbox.Ensure(ctx, wd, sandbox.ExtraWorkspace(wd, agentRef), template, configDir)
+	if err != nil {
+		return err
+	}
+
+	// Resolve env vars the agent needs and forward them into the sandbox.
+	// Docker Desktop proxies well-known API keys automatically; this handles
+	// any additional vars (e.g. MCP tool secrets).
+	envFlags, envVars := sandbox.EnvForAgent(ctx, agentRef, environment.NewDefaultProvider())
+
+	// Forward the gateway as an env var so docker sandbox exec sets it
+	// directly inside the sandbox.
+	if gateway := runConfig.ModelsGateway; gateway != "" {
+		envFlags = append(envFlags, "-e", envModelsGateway+"="+gateway)
+	}
+
+	dockerCmd := sandbox.BuildExecCmd(ctx, name, cagentArgs, envFlags, envVars)
+	slog.Debug("Executing in sandbox", "name", name, "args", dockerCmd.Args)
+
+	if err := dockerCmd.Run(); err != nil {
+		if exitErr, ok := errors.AsType[*exec.ExitError](err); ok {
+			// Clean up the token writer before exiting so the temp
+			// file is removed (defers won't run after os.Exit).
+			stopTokenWriter()
+			os.Exit(exitErr.ExitCode()) //nolint:gocritic // intentional exit to propagate sandbox exit code
+		}
+		return fmt.Errorf("docker sandbox exec failed: %w", err)
+	}
+
+	return nil
+}

docs/configuration/agents/index.md

@@ -232,10 +232,6 @@ agents:
     toolsets:
       - type: filesystem
       - type: shell
-        sandbox:
-          image: golang:1.23-alpine
-          paths:
-            - "."
       - type: think
       - type: todo
 

docs/configuration/overview/index.md

@@ -124,7 +124,7 @@ Models can be referenced inline or defined in the `models` section:
   <a class="card" href="{{ '/configuration/sandbox/' | relative_url }}">
     <div class="card-icon">📦</div>
     <h3>Sandbox Mode</h3>
-    <p>Run shell commands in an isolated Docker container for security.</p>
+    <p>Run agents in an isolated Docker container for security.</p>
   </a>
   <a class="card" href="{{ '/configuration/structured-output/' | relative_url }}">
     <div class="card-icon">📋</div>

docs/configuration/sandbox/index.md

@@ -1,16 +1,16 @@
 ---
 title: "Sandbox Mode"
-description: "Run shell commands in an isolated Docker container for enhanced security."
+description: "Run agents in an isolated Docker container for enhanced security."
 permalink: /configuration/sandbox/
 ---
 
 # Sandbox Mode
 
-_Run shell commands in an isolated Docker container for enhanced security._
+_Run agents in an isolated Docker container for enhanced security._
 
 ## Overview
 
-Sandbox mode runs shell tool commands inside a Docker container instead of directly on the host system. This provides an additional layer of isolation, limiting the potential impact of unintended or malicious commands.
+Sandbox mode runs the entire agent inside a Docker container instead of directly on the host system. This provides an additional layer of isolation, limiting the potential impact of unintended or malicious commands.
 
 <div class="callout callout-info">
 <div class="callout-title">ℹ️ Requirements
@@ -19,111 +19,44 @@ Sandbox mode runs shell tool commands inside a Docker container instead of direc
 
 </div>
 
-## Configuration
+## Usage
+
+Enable sandbox mode with the `--sandbox` flag on the `docker agent run` command:
+
+```bash
+docker agent run --sandbox agent.yaml
+```
+
+This runs the agent inside a Docker container with the current working directory mounted.
+
+## Example
 
 ```yaml
+# agent.yaml
 agents:
   root:
     model: openai/gpt-4o
     description: Agent with sandboxed shell
     instruction: You are a helpful assistant.
     toolsets:
       - type: shell
-        sandbox:
-          image: alpine:latest # Docker image to use
-          paths: # Directories to mount
-            - "." # Current directory (read-write)
-            - "/data:ro" # Read-only mount
 ```
 
-## Properties
-
-| Property | Type   | Default         | Description                                   |
-| -------- | ------ | --------------- | --------------------------------------------- |
-| `image`  | string | `alpine:latest` | Docker image to use for the sandbox container |
-| `paths`  | array  | `[]`            | Host paths to mount into the container        |
-
-## Path Mounting
-
-Paths can be specified with optional access modes:
-
-| Format       | Description                                     |
-| ------------ | ----------------------------------------------- |
-| `/path`      | Mount with read-write access (default)          |
-| `/path:rw`   | Explicitly read-write                           |
-| `/path:ro`   | Read-only mount                                 |
-| `.`          | Current working directory                       |
-| `./relative` | Relative path (resolved from working directory) |
-
-Paths are mounted at the same location inside the container as on the host, so file paths in commands work the same way.
-
-## Example: Development Agent
-
-```yaml
-agents:
-  developer:
-    model: anthropic/claude-sonnet-4-0
-    description: Development agent with sandboxed shell
-    instruction: |
-      You are a software developer. Use the shell tool to run
-      build commands and tests. Your shell runs in a sandbox.
-    toolsets:
-      - type: shell
-        sandbox:
-          image: node:20-alpine # Node.js environment
-          paths:
-            - "." # Project directory
-            - "/tmp:rw" # Temp directory for builds
-      - type: filesystem
+```bash
+docker agent run --sandbox agent.yaml
 ```
 
 ## How It Works
 
-1. When the agent first uses the shell tool, docker-agent starts a Docker container
-2. The container runs with the specified image and mounted paths
-3. Shell commands execute inside the container via `docker exec`
-4. The container persists for the session (commands share state)
-5. When the session ends, the container is automatically stopped and removed
-
-## Container Configuration
-
-Sandbox containers are started with these Docker options:
-
-- `--rm` — Automatically remove when stopped
-- `--init` — Use init process for proper signal handling
-- `--network host` — Share host network (commands can access network)
-- Environment variables from host are forwarded to container
-
-## Orphan Container Cleanup
-
-If docker-agent crashes or is killed, sandbox containers may be left running. docker-agent automatically cleans up orphaned containers from previous runs when it starts. Containers are identified by labels and the PID of the docker-agent process that created them.
-
-## Choosing an Image
-
-Select a Docker image that has the tools your agent needs:
-
-| Use Case               | Suggested Image      |
-| ---------------------- | -------------------- |
-| General scripting      | `alpine:latest`      |
-| Node.js development    | `node:20-alpine`     |
-| Python development     | `python:3.12-alpine` |
-| Go development         | `golang:1.23-alpine` |
-| Full Linux environment | `ubuntu:24.04`       |
-
-<div class="callout callout-tip">
-<div class="callout-title">💡 Custom Images
-</div>
-  <p>For complex setups, build a custom Docker image with all required tools pre-installed. This avoids installation time during agent execution.</p>
-
-</div>
+1. When `--sandbox` is specified, docker-agent launches a Docker container
+2. The current working directory is mounted into the container
+3. All agent tools (shell, filesystem, etc.) operate inside the container
+4. When the session ends, the container is automatically stopped and removed
 
 <div class="callout callout-warning">
 <div class="callout-title">⚠️ Limitations
 </div>
 
-- Only the <code>shell</code> tool runs in the sandbox; other tools (filesystem, MCP) run on the host
-- Host network access means network-based attacks are still possible
-- Mounted paths are accessible according to their access mode
 - Container starts fresh each session (no persistence between sessions)
 
 </div>
@@ -140,10 +73,6 @@ agents:
     instruction: You are a helpful assistant.
     toolsets:
       - type: shell
-        sandbox:
-          image: node:20-alpine
-          paths:
-            - ".:rw"
       - type: filesystem
 
 permissions:

docs/configuration/tools/index.md

@@ -66,7 +66,6 @@ The agent has access to the full system shell and environment variables. Command
 | Property  | Type   | Description                                                                      |
 | --------- | ------ | -------------------------------------------------------------------------------- |
 | `env`     | object | Environment variables to set for all shell commands                              |
-| `sandbox` | object | Run commands in a Docker container. See [Sandbox Mode](/configuration/sandbox/). |
 
 ### Think