Nither

Nither is a local-first AI agent and coding CLI for Windows and other supported platforms.

• Native llama.cpp runtime — no Ollama dependency
• Qwen3.5-4B GGUF local-first workflow
• Streaming responses with a live thinking indicator
• Native model manager: /download, /remove, /update, /models, /reinstall-qwen
• Real model download progress with size remaining and ETA
• Hardware detection and automatic model selection
• Persistent context and memory
• Generic MCP support
• Multi-agent architecture
• Project and Git intelligence
• Desktop file/folder creation for agent-built deliverables
• Multi-file project generation with fresh local-model context per file
• GPU device-loss recovery with CPU fallback during project generation
• Project verification: missing/empty files and Python syntax checks
• Automatic repair of failed generated files, with bounded repair rounds
• Dedicated Snake & Ladder project build plan
• Windows installation support
• Automatic Nither update checking
• Safe CPU fallback for Windows Vulkan/device-loss problems
• Fast background downloads and update checks

See PHASE6_RELEASE.md for the Phase 6 release details and validation information.

Install llama.cpp if needed:

winget install llama.cpp

Install Nither once from the source checkout:

python -m pip install -e .

Then simply run:

nither

Future Nither updates can be checked automatically at startup; you do not need to repeatedly run the editable install command for normal updates.

openresearchtools/Qwen3.5-4B-GGUF:Q4_K_M

Nither manages the local model and uses llama.cpp directly. No Ollama installation or API key is required for the default local workflow.

/help
/models
/download <model>
/remove <model>
/update <model>
/reinstall-qwen
/runtime
/hardware
/auto-model
/test
/code
/agents
/mcp
/memory

For a request such as Build a calculator, Nither can create the requested source file on the user's Desktop using its bounded filesystem tools.

Run the test suite with:

python -m pytest -q

Phase 6 validation: 94 tests passed, Python compilation passed, and wheel build passed.