Skip to main content

Interactive Examples

agentvisor exec mode bridges your terminal directly into the sandbox, letting you run any interactive command — Claude Code, Aider, a shell, or a custom script — with the same policy-enforced HTTP proxying and credential substitution that serve mode provides.

How It Differs from serve/run

Aspectagentvisor serveagentvisor exec
Entry pointLangGraph graph via HTTP APIAny command (shell, AI coding tool, etc.)
Temporal requiredYesNo
HTTP API exposedYes (port 8090)No
Session typeNon-interactive, API-drivenInteractive, TTY-attached
CheckpointingYes (Temporal)No
SSE streamingYesNo

When to Use exec Mode

  • Running Claude Code or another AI coding assistant inside an isolated sandbox
  • Running Aider, Cursor (CLI), or similar tools that need network access restricted to specific APIs
  • Dropping into a bash shell inside the sandbox for debugging or exploration
  • Executing a custom one-shot script that needs the proxy or MCP gateway but not Temporal

Prerequisites

  • No Temporal required
  • No HTTP API configuration needed
  • Just agentvisor exec . (or --command to override)

The framework configuration (mav-agent-config.yaml) selects the interactive provider and uses framework.command to specify what runs inside the sandbox:

framework:
provider: interactive
command: ["claude", "--dangerously-skip-permissions"]

Feature Availability

FeatureAvailable in exec mode
HTTP proxy (TLS termination + policy)Yes
Credential substitutionYes
MCP gatewayYes
A2A gatewayYes
Key-value storeYes
Log forwardingYes
CheckpointingNo — requires Temporal
SSE streamingNo — requires HTTP API

Examples

ExampleDescription
Simple ShellThe minimum interactive provider config — a plain /bin/sh shell with an Ollama-only HTTP policy, no credential brokering
Claude CodeSandboxed Claude Code CLI with credential brokering and Anthropic-only HTTP policy
VS Code TunnelFull VS Code IDE connected via Microsoft's tunnel relay, sandboxed with a GitHub + VS Code infrastructure HTTP policy

See Also