Traditional AI chat generates text. PyShell generates and executes Python code in a persistent runtime, using execution results to drive the conversation.
- 🧠 Persistent sessions: Variables, imports, and execution context carry over between tasks.
- 🔌 No MCP servers needed: Install Python libraries and let the agent discover and use their APIs directly.
- 🔄 Execution feedback: The agent observes output and errors, correcting mistakes and iterating automatically.
- 📊 Interactive visualizations: The agent creates and updates charts, graphs, and maps directly in your browser.
- 🎙️ Voice-driven: A voice agent understands your requests and delegates execution to a Python worker.
demo.mp4
Accepted adapter: LongLink/pyshell-worker.
Base model: Qwen/Qwen3.5-9B.
Task success: 58.33% (35/60 tasks). Valid Python: 99.38% (159/160 responses).
Evaluation date (UTC): 2026-11-11T15:01:13.152443+00:00. Evaluation workflow.
Set OPENAI_API_KEY in sample/.env, then run from the repository root:
make upOpen http://127.0.0.1:8010 in your browser.
- Speak: Give instructions to the voice agent in natural language.
- Delegate: The voice agent understands your request and passes the task to a Python worker.
- Execute & iterate: The worker discovers available Python libraries, writes and executes code, and uses output and errors to refine its actions.
- Visualize: Results appear as interactive charts and visualizations directly in your browser.
- Persist: Variables, data, and execution context remain available for follow-up tasks.
Run from the repository root:
make up # Setup the local server
make down # Stop and remove the containers
make test # Testing
make format # Format the code
make logs # Show the container logsLongLink 2026

