Reika

Reika is a coding agent that runs in your terminal and works with any OpenAI-compatible model server: llama.cpp, MLX, vLLM, or a cloud API. It was designed around small local models first (8B–35B, often heavily quantized, on a 16–32k context window), so it is careful with context, fails more gracefully, and says when it's stuck. It doesn't make a small model smarter. It makes working with one less frustrating.

Highlights:

• Context discipline: old tool output collapses to short summaries, and the engine's prompt cache is preserved between turns. When the window fills, the model writes its own findings note before older turns are folded away.
• Loop and spiral breaking: repeated reads and re-derived reasoning are detected and answered with an escalating response that ends in an honest stop, not a 30-minute spiral.
• Checks on the model's work: blind edits are sent back for a read first, TypeScript edits are typechecked, and written plans are tracked step by step.
• Plan mode: read-only exploration that ends in a file-specific plan you can refine before running it.
• Safe by default: dangerous commands prompt, and on macOS shell commands run in a kernel sandbox.
• Extensible with MCP servers and Markdown skills.
• No telemetry: with a local model, your code never leaves your machine.

Free and open source (Apache-2.0). macOS and Linux, Node.js 22+.