CLI Agent: Subconscious + E2B

A developer-first autonomous agent that reasons and executes code in a secure cloud sandbox. This agent uses Subconscious for long-horizon reasoning and E2B for isolated code execution.

  ┌─┐┬ ┬┌┐ ┌─┐┌─┐┌┐┌┌─┐┌─┐┬┌─┐┬ ┬┌─┐
  └─┐│ │├┴┐│  │ ││││└─┐│  ││ ││ │└─┐
  └─┘└─┘└─┘└─┘└─┘┘└┘└─┘└─┘┴└─┘└─┘└─┘  + E2B Sandbox

What it does

• Long-horizon reasoning: Subconscious handles complex multi-step tasks with planning and self-correction
• Secure execution: All code runs in isolated E2B cloud sandboxes
• File I/O: Upload local files, download generated outputs (charts, reports, data)
• Multi-language: Python, JavaScript, TypeScript, Go, Rust, C++, Ruby, Java, Bash
• Data science ready: numpy, pandas, matplotlib pre-installed
• Command history: Arrow keys navigate previous commands
• Zero-config tunneling: Works out of the box - no external dependencies

Quick Start

Option 1: Use npx (easiest)

npx @subcon/e2b-cli

The CLI will prompt you for API keys on first run and save them for future sessions.

Option 2: Install globally

npm install -g @subcon/e2b-cli
e2b-cli

Option 3: Run from source (for development)

# Install dependencies
bun install

# Set your API keys
export SUBCONSCIOUS_API_KEY=your_key_here  # Get at https://subconscious.dev/platform
export E2B_API_KEY=your_key_here           # Get at https://e2b.dev

# Run the agent
bun run agent

Example Usage

Try it with the included demo file

▸ Task › Analyze file: ./demo_data.csv and create a summary of hours worked per employee
▸ Context ›

Simple task (no files)

▸ Task › Calculate the first 50 Fibonacci numbers and identify which ones are prime
▸ Context ›

Generate a chart

▸ Task › Analyze file: ./demo_data.csv and create a bar chart of total billable hours by department. Save to output: ./chart.png
▸ Context › Use pandas and matplotlib

Full analysis with multiple outputs

▸ Task › Analyze file: ./demo_data.csv and do the following:
1. Create a chart of hours by employee. Save to output: ./hours_chart.png
2. Write a markdown report with key insights. Save to output: ./analysis.md
▸ Context › Use pandas and matplotlib

File Handling

The agent supports file upload and download:

| Syntax | Description | Example | |--------|-------------|---------| | file: ./path | Upload a file to the sandbox | Analyze file: ./data.csv | | '/path/to/file' | Quoted path (drag-and-drop) | Analyze '/Users/me/data.csv' | | files: ./dir/*.csv | Upload multiple files (glob) | Process files: ./reports/*.csv | | output: ./path | Download output when done | Save chart to output: ./chart.png |

Tip: You can drag and drop files into the terminal and the quoted path will be automatically recognized and uploaded to the sandbox.

Supported output formats:

• Images: .png, .jpg, .gif, .webp, .svg
• Documents: .md, .txt, .html, .pdf
• Data: .json, .csv, .xml
• Archives: .zip, .tar, .gz

How it works

1. Input files are uploaded to /home/user/input/ in the sandbox
2. Output files are saved to /home/user/output/ in the sandbox
3. When the agent completes, outputs are automatically downloaded to your specified local paths

Supported Languages

| Language | Runtime | Notes | |----------|---------|-------| | Python | python3 | numpy, pandas, matplotlib pre-installed | | JavaScript | node | Node.js runtime | | TypeScript | ts-node | Via npx | | Bash | bash | Shell scripts | | Go | go run | Single file execution | | Rust | rustc | Compiled before execution | | C++ | g++ | Compiled with -O2 | | C | gcc | Compiled with -O2 | | Ruby | ruby | Standard runtime | | Java | javac + java | Compiled and run |

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                         CLI Agent                                │
├─────────────────────────────────────────────────────────────────┤
│  User Input  ->  File Parser  ->  Subconscious API  ->  Display │
│                      |                  |                        │
│               Upload Files      Stream Reasoning                 │
│                      |                  |                        │
│              E2B Sandbox  <---  Tool Calls (execute_code)       │
│                      |                                           │
│              Download Outputs  ->  Local Filesystem              │
└─────────────────────────────────────────────────────────────────┘

Source Structure

src/
  cli/
    run.ts              # Interactive CLI with command history
    fileParser.ts       # Parses file:/output: references
  e2b/
    sandbox.ts          # E2B sandbox wrapper with multi-language support
  tools/
    e2bServer.ts        # HTTP server exposing execute_code tool
    tunnel.ts           # Tunnel management (localtunnel)
  types/
    agent.ts            # TypeScript definitions
  config.ts             # Configuration loading
  index.ts              # Entry point

Configuration

Environment Variables

| Variable | Required | Description | |----------|----------|-------------| | SUBCONSCIOUS_API_KEY | Yes | Your Subconscious API key | | E2B_API_KEY | Yes | Your E2B API key | | TUNNEL_URL | No | Use existing tunnel instead of auto-start | | VERBOSE | No | Enable verbose logging (true/false) |

Config file (optional)

Create agent.config.json to customize:

{
  "tunnel": {
    "enabled": true,
    "autoStart": true,
    "port": 3001
  },
  "tools": {
    "port": 3001,
    "host": "localhost"
  },
  "environment": {
    "filterSensitive": true,
    "sensitivePatterns": ["KEY", "SECRET", "TOKEN", "PASSWORD"]
  }
}

Troubleshooting

Tunnel not starting

The agent uses localtunnel which is installed automatically with bun install. If you're having issues:

1. Check your network - localtunnel requires outbound internet access
2. Use an existing tunnel - Set TUNNEL_URL environment variable to bypass auto-start

Output files not appearing

1. Make sure you use the output: prefix in your task
2. Check that the agent says [file] Downloaded: at the end
3. Verify the path is accessible (not a protected system directory)

Sandbox timeout or slow startup

The first run may take 30-60 seconds as E2B provisions the sandbox and installs packages. Subsequent runs are faster.

Links

• Subconscious - Long-horizon AI reasoning
• E2B - Secure code sandboxes
• localtunnel - Tunnel service

License

Apache-2.0