Overview
PeaBox LLM Workbench is a local-first, multi-provider AI workspace built with Streamlit. It provides a single interface for working with Ollama, OpenAI, Anthropic, Google Gemini, Groq, OpenRouter, LiteLLM, LM Studio and other OpenAI-compatible endpoints.
Instead of switching between different AI applications, provider dashboards, and local model interfaces, PeaBox gives you one workspace for conversations, model selection, files, vision inputs, reasoning controls and generation settings.
Features
- Multi-provider support: Ollama (local), SSH Ollama over SSH, OpenAI, Anthropic, Google Gemini, Groq, OpenRouter, Mistral, Together AI, custom OpenAI-compatible endpoints
- Local conversation history stored in SQLite
- Persistent conversations across restarts (history snapshot before changes)
- Conversation titles auto-generated, renaming, pinning, archiving and deleting
- Export conversations as JSON
- File attachments: PDF, DOCX, PPTX, TXT, Markdown, XML, HTML, YAML, MLY
- Spreadsheets and data: XLSX, XLSM, CSV, JSON
- Code uploads: Python, JavaScript, TypeScript, PowerShell, HTML, config files, logs
- Vision support: images sent to compatible vision models (PNG, JPG, JPEG, WEBP)
- Reasoning controls: Default / Off / Low / Medium / High / Max
- Generation controls: Temperature, Top P, Top K, Repeat Penalty, Context Tokens, Maximum Output Tokens, Reasoning Level, "Use Model Defaults"
- Settings & API Vault: separate provider configuration (Base URL, API Key, Default Model, SSH Host/User/Port/Key)
Model Discovery
From the Workbench, select Discover / Refresh Models. The application retrieves available models exposed by supported providers and makes them available for selection without changing applications.
Especially useful when experimenting with local models because your installed model collection may change frequently.
Conversations
Conversations are stored locally using SQLite. You can create new ones, view history, reopen previous conversations, archive, delete and export as JSON. Conversation history remains after restarting the app.
Working With Documents
Upload supported documents (PDF, DOCX, PPTX, TXT, Markdown, XML, HTML, YAML). Ask questions such as "Summarize these documents", "Compare these two files", or "Identify the differences". Document content is added to the model context, subject to extraction limits.
Working With Code
Upload source files and ask the selected model to explain code, find bugs, refactor functions, generate unit tests, convert implementations to another language, or produce analysis scripts.
Note: PeaBox does not automatically execute generated code. Run `codeproof scan .` or use the CLI to verify.
Security Considerations
- Keep API keys out of source files; never commit credentials, SSH private keys, passwords or .env files.
- Never add SSH private keys to repositories.
- `.gitignore`: .venv/, __pycache__/, *.pyc, .env, .peabox/, .peabox_streamlit_llm/, .streamlit/secrets.toml, *.key, *.pem.
Privacy Model
Using PeaBox as a local interface does not make an external cloud model local. Privacy and data-handling policies of the selected model provider still apply. Local runs process image bytes on your computer only.
Troubleshooting
- Ollama models don't appear: check if Ollama is running and verify base URL (http://localhost:11434).
- Remote Ollama connection fails: verify SSH host/user/port/private key path.
- Cloud provider connection fails: verify API key, base URL, model name, network connectivity.
- Python dependency problems: upgrade pip and reinstall requirements.txt.
Product Roadmap
v4.2 Chat experience (streaming responses, regenerate, side-by-side model comparison). v4.3 Knowledge & RAG (workspaces, local document index, cited answers, prompt library). v4.4 Agents (tool calling, MCP support, human approval, provider fallback). v4.5 Observability (usage/cost analytics, budget caps, backups). v5.0 Teams (authentication, shared workspaces, audit logs, SSO).