Agent Chat

A hackable ChatGPT-like code interpreter for data analysis, code execution, and image generation. Real-time token streaming, inline tool execution, and steering/abort controls.

Built with liteagent + FastHTML + DaisyUI + HTMX + Modal.

Features

• Secure cloud sandboxes — Python runs in isolated Modal containers, fully sandboxed from the host
• Persistent state — Variables, imports, and definitions carry over between code executions
• Vision-enabled tool results — The agent sees stdout, stderr, and any generated images (plots, charts, etc.)
• Auto-captured plots — Matplotlib/seaborn figures are captured as images; Plotly figures are captured as interactive HTML for you to explore
• Interactive Plotly charts — Plotly figures render as fully interactive charts you can zoom, pan, and hover
• Gemini image generation — Generate and edit images using Google's Gemini API directly in code
• Auto-captured PIL images — Any PIL Image assigned to a variable is automatically captured and shown
• Data science ready — pandas, numpy, scipy, scikit-learn, matplotlib, seaborn, plotly, and more pre-installed; install any package with pip
• Per-user isolation — Each browser session gets its own sandbox; conversation and state persist across page refreshes and new tabs
• Token-by-token streaming — Response text, thinking, and tool call arguments stream in real-time via SSE
• Steering and abort — Send follow-up messages while the agent is running to steer it, or stop it mid-run
• Extended thinking — See the model's reasoning process inline (when supported)
• Multi-provider LLM — Uses liteagent + LiteLLM for easy model switching

Setup

uv sync
cp .env.example plash.env  # Then edit plash.env with your keys

Note: This project uses plash.env (not .env) so it works seamlessly with Plash deployment.

Environment Variables

# Required for code execution (get from https://modal.com)
MODAL_TOKEN_ID=
MODAL_TOKEN_SECRET=

# Required for session encryption (generate any random string)
FAST_APP_SECRET=

# LLM API keys — only add the one you plan to use
# Model is set in agents/tools.py get_agent() — change the model= line to switch
ANTHROPIC_API_KEY=   # For Claude models
OPENAI_API_KEY=      # For GPT models (optional)
GOOGLE_API_KEY=      # For Gemini models and image generation (optional)

Development

# Run the app locally
uv run python main.py

# Run tests (skips slow integration tests by default)
./dev test

# Run only slow integration tests (hits real APIs)
./dev test -m slow

# Run ALL tests (including slow)
./dev test -m ""

# Run tests with coverage
./dev test -m "" --cov=agents --cov=main --cov-report=term-missing

# Lint and format
./dev lint

Project Structure

agents/
  tools.py           # Per-user Agent management, tool factory, sandbox lifecycle
  prompts.py         # System prompt generation
  coding_sandbox.py  # Modal sandbox wrapper for code execution
  driver_program.py  # Runs inside Modal sandbox, executes code
  ui/
    components.py    # Single-stream UI components and SSE event renderer
    markdown.py      # Markdown rendering with syntax highlighting
    tool_renderers.py # Custom tool call display
tests/               # pytest tests
main.py              # FastHTML app, routes, and SSE bridge

Export requirements.txt

uv export --no-hashes --no-dev -o requirements.txt

Deploy to Plash

uv run plash_deploy