DryDock — Local-first CLI coding agent

What it is

DryDock is a TUI coding assistant designed to work with local LLMs. It gives you a conversational interface to your codebase — explore, modify, build, and test projects through natural language and a focused set of tools.

No data leaves your machine. No API keys. No per-token billing. Just your laptop, your model, and your code.

Install

From PyPI:

# Python 3.12+
pip install drydock-cli
drydock

From Docker Hub:

docker run -it --rm \
    -e LLAMACPP_URL=http://host.docker.internal:8001/v1 \
    -e LLAMACPP_MODEL=gemma4 \
    --add-host=host.docker.internal:host-gateway \
    -v "$HOME/.drydock:/root/.drydock" \
    -v "$(pwd):/work" \
    fbobe3/drydock:latest

From source:

git clone https://github.com/fbobe321/drydock.git
cd drydock && pip install -e .

Recommended serving stack

DryDock is tested and optimized for Gemma 4 26B-A4B served by llama.cpp with --jinja (the chat-template fix that prevents tool-call loops). Other OpenAI-compatible providers (Ollama, LM Studio, Mistral, OpenAI, Anthropic) work but are not as thoroughly tested.

# 1. Download Unsloth's GGUF
huggingface-cli download unsloth/gemma-4-26B-A4B-it-GGUF \
    --include "gemma-4-26B-A4B-it-UD-Q3_K_M.gguf" \
    --local-dir /path/to/models

# 2. Start llama-server with --jinja
./build/bin/llama-server \
    -m /path/to/models/gemma-4-26B-A4B-it-UD-Q3_K_M.gguf \
    --host 0.0.0.0 --port 8000 \
    -ngl 99 -c 32768 -np 1 \
    --jinja -ctk q8_0 -ctv q8_0 \
    --alias gemma4

See DEPLOYMENT.md for the full ~/.drydock/config.toml, env vars, and troubleshooting.

What's in the box

⎈TUI-first

Textual-powered terminal UI. Slash commands, plan/edit modes, session history, undo/back/goal.

⚒14 built-in tools

read_file, write_file, search_replace, bash, grep, glob, mechanical_rename, retrieve, and more.

⚓Local-first

OpenAI-compatible endpoint. Works with llama.cpp, Ollama, LM Studio, vLLM. No cloud required.

⌖GraphRAG

AST symbol indexer + TF-IDF retriever. Auto-prefetches relevant code on every turn.

⚙Adaptive thinking

Per-turn thinking budget: HIGH for planning, OFF for routine writes, LOW for recovery.

⌗Multimodal vision

Optional mmproj-F16.gguf projector enables OpenAI-style image inputs with Gemma 4.

Tested hardware

GPU	2× NVIDIA RTX 4060 Ti 16GB (32GB total VRAM)
RAM	64GB recommended, 32GB minimum
Model	Gemma 4 26B-A4B (26B MoE, 4B active params per token)
Performance	~15–17 tok/s decode (llama.cpp Q3_K_M)
OS	Ubuntu 22.04 / 24.04, kernel 6.8+
Min single-GPU	24GB VRAM card with reduced context