Claude Code In a Box

I want to think of Claude Code itself as a deployable general agent.

Dot Claude

Claude Code, as the harness, can run any combination of commands, agents, skills, output-styles, and hooks, until it returns the desired output, in any format.

The instructions that drive this output, written mostly in natural language, reside in the .claude/ folder. This can be viewed as a complete, portable agent definition.

.claude/
├── settings.json      # Model, permissions, hooks, output-style
├── commands/          # Slash commands
├── agents/            # Subagent definitions
├── hooks/             # Event triggers
├── output-styles/     # Persona overrides, generalization
└── skills/
    └── any-skill/       # Behaviors, workflows
        ├── SKILL.md       # Instructions, workflows
        ├── references/    # Context
        ├── templates/     # Response templates
        └── scripts/       # Scripts, tools

This means you can draft, iterate, and test locally on your machine. Once it is stable, you can structure a simple workflow like this:

workflow/
├── .claude/ # complex tasks to be performed and verified
├── inputs/  # request, data
├── outputs/ # response
└── CLAUDE.md # guidance, memory

You can run the agent something like this:

claude -p "follow these instructions"  

Which are of course all contained in the .claude/. Through prompt chaining, tool calling, progressive disclosure (as sequenced instructions, in, for example, a Skill), or as asynchronous, multi-threaded processes (parallel agents), Claude Code completes the work and writes it to outputs/.

Think of Docker as an analogy (however imprecise and flawed): As your agent runs locally, so it will in the cloud.

Following that analogy, once a workflow is stable and verified, you can deploy Claude Code + ./claude/ in a wrapper on a virtual machine, a sandbox. A box.

Box

In the example that follows, the box runs on an e2b sandbox,¹ much like "Claude on the Web" runs, but with a custom .claude/.

All of the prompts, context, and tools reside in the .claude/ folder, so we can simply deploy Claude Code as the application, here with a React UI, a thin FastAPI layer, and a BOX – a remote sandbox running Claude Code with a copied .claude/ folder.

Example

Mortals is a character generator for the game Cupid.

To generate characters for this game, I developed a system of sequenced prompts (progressive disclosure) and tools (uv in-line scripts) into a workflow: Claude Code generates birthdate information at random, with which it produces an astrological natal chart with SVG. It then writes a complete biography of the character based on the chart, and generates a portrait with Nano Banana.

Like this:

This is a sped-up screen recording of a roughly 2 minute process using this .claude/:

When you initiate a session, the server uploads the .claude/ folder to the e2b (box) instance, manages its lifecycle, pipes the streams to the UI, and logs the session as files.² Webhooks trigger UI state changes when the files save to the box.

This makes the application scalable, with on-demand cost.³

Change the .claude/ folder, and the application assumes a completely different role, function, skills. Same infrastructure.

What Makes This Distinct

(assessment by Opus 4.5)

Existing implementations are infrastructure-focused ("run Claude Code remotely"). This framing is architectural: applications expressed in natural language via .claude/ folders. It's a development paradigm, not just a deployment pattern.