Installation harvey
harvey is a terminal coding agent that runs language models locally. No cloud account or API key required.
Requirements
Hardware requirements
Harvey runs entirely on local hardware, so the practical limits are RAM (to hold the model) and storage (for the model files themselves), not just raw compute.
- RAM — 8 GB minimum for models up to ~7B parameters
(matches the default
Qwen2.5-Coder-7Brecommendation below); 16 GB or more gives headroom for larger models, or lets a 7B-class model run while leaving RAM free for the rest of your work. - Storage — at least 64 GB free if you expect to try
more than one model. Individual llamafiles and Ollama models routinely
run several GB each (the
Qwen2.5-Coder-7B-Q5_K_S.llamafilebelow is ~5 GB), and Harvey’s RAG stores and session recordings add up over time. - Raspberry Pi — a Raspberry Pi 5 (8 GB or more) runs Harvey fine with CPU-only inference; expect roughly 2–10 tokens/s depending on model size and quantization without dedicated AI accelerator hardware. A hardware NPU accelerator (e.g. Raspberry Pi’s AI HAT+ 2) is not currently supported as a backend.
Compile from source
git clone https://github.com/rsdoiel/harvey
cd harvey
make
make test
make installmake install copies the binaries into
$HOME/bin/ by default. Override with
make install prefix=/usr/local.
Model backend
Ollama (recommended)
Install Ollama from https://ollama.com/download, then pull a model:
ollama pull qwen2.5-coder:7bHarvey detects Ollama automatically on startup.
Llamafile (self-contained, no install)
Download a pre-built llamafile from: https://docs.mozilla.ai/llamafile/getting-started/pre-built-llamafiles
Recommended: - Qwen2.5-Coder-7B-Q5_K_S.llamafile (~5 GB,
good for most hardware) -
Phi-3.5-mini-instruct-Q4_K_M.llamafile (~2 GB, low-VRAM /
CPU)
Place it in ~/Models/ and make it executable (Linux /
macOS):
chmod +x ~/Models/Qwen2.5-Coder-7B-Q5_K_S.llamafileHarvey finds the llamafile automatically and connects.
Running harvey
cd ~/myproject
harvey