cat67.ai — Local AI Development Workspace
A private desktop workspace for AI-assisted software development that keeps projects, conversations, task state, and model inference on the developer’s own computer.
Overview
cat67.ai is a private desktop workspace for developers who want AI-assisted coding without sending their projects, conversation history, or task state to a cloud service. It combines a local language model with project-aware tools, structured context management, and a safety layer for reviewing changes before they are applied.
Features
- Project understanding through files, search, Git, notebooks, PDFs, and images
- Auditable tool workflows with arguments, results, durations, and status
- Permission modes, exact-operation approvals, diffs, checkpoints, and undo
- Visible context budgeting with deterministic compression before overflow
- Structured Ultra plans and sequential specialist sub-agents
- Image and file attachments with local OCR and structured fallbacks
- Application inspection through a View App tool
- Local diagnostics with live RAM and VRAM information
Problem
Many AI coding workflows require source code and developer context to leave the machine. Developers also need a clear record of tool use, controlled changes, and enough context continuity to complete longer tasks without making the model context unmanageable.
Solution
The workspace keeps projects, chats, tool history, task state, checkpoints, diagnostics, and semantic memory on the developer’s computer while running the selected model locally. It separates full interface history from active model context, persistent task state, semantic project memory, and a file-context cache.
Architecture
Desktop: Electron interface for projects, chats, file preview, tool history, terminal, and diagnostics.
Control plane: A local FastAPI service manages the agent loop, prompt and context construction, permissions, checkpoints, SQLite application state, and Qdrant semantic memory.
Inference: One managed llama.cpp runtime serves the selected on-device model, including Qwen 3.5 9B Q8_0 and GPT-oss 20B F16.
Results
The resulting application supports end-to-end local coding workflows with project-aware tools, persistent state, explicit review controls, local inference, and automated validation through Python unit and integration tests, Playwright Electron end-to-end tests, and release-performance benchmarks.
Challenges
The central engineering challenges were maintaining useful context within bounded model windows, preserving a reviewable task history, safely resolving file operations inside a canonical project root, coordinating specialist agents through one inference runtime, and enforcing clear security boundaries around tools and local services.