100% OFFLINE • ZERO API KEYS • LOCAL WEIGHTS ONLY

Agentic terminal coding,
powered by your own hardware.

A fast, keyboard-driven terminal coding tool built for local models. Run DeepSeek, Qwen, Llama, and Gemma directly on your machine with Ollama and zero cloud telemetry.

git clone https://github.com
View on GitHub ↗
deepcode — ~/projects/Deepcode (main)
ollama :: deepseek-coder:6.7b
❯ refactor auth middleware to support local Ed25519 keys and add unit tests
✻ Thinking locally (48.2 tok/s • 14.1 GB VRAM)...
⏺ Read(src/middleware/auth.ts)
⎿ Read 142 lines
⏺ Edit(src/middleware/auth.ts)
- import { verifyJwt } from '../legacy/jwt';
+ import { verifyEd25519 } from '../crypto/ed25519';
+ const pubKey = await loadLocalKey('.deepcode/id_ed25519.pub');
⏺ Bash(npm test src/middleware/auth.test.ts)
⎿ PASS src/middleware/auth.test.ts (6 passed, 0 failed in 0.42s)
Updated auth.ts to verify Ed25519 signatures locally and added 6 test cases covering key rotation and invalid signatures. No external dependencies added.
❯ Ask deepcode to edit files, run tests, or inspect git diff...

Built for air-gapped speed & privacy

🔒

Zero Cloud Lock-In

No accounts, no rate limits, and no source code leaving localhost. Safe for proprietary repositories and air-gapped workstations.

⚡

Native Tool Calling

Structured file diffs, ripgrep search, shell execution, and git staging tuned specifically for open-weight coding models.

🧠

Smart Context Pruning

Automatically compresses AST outlines and git diffs so 8B–32B local models stay sharp inside 32k–128k context windows.

🎛️

Hot-Swap Models

Use a fast 8B model for instant file edits and auto-checks, then switch to a 32B+ reasoning model for architecture refactors.

Works with your local inference engine

Connects out of the box to Ollama and local servers.

Ollama DeepSeek llama.cpp LM Studio

Get the code & setup instructions on GitHub ↗