Harvey
Harvey is an agent REPL written in Go and designed to use an Ollama (or hailo-ollama) server to access language models locally. It is a terminal based application.
The Harvey name was inspired by the play of that name by Mary Chase. I saw parallels between the story Harvey and my personal language model agent. Many people think of agents only in the context of very big companies. I think small models running on small or tiny computers are an opportunity. Harvey, is a small agent for small and tiny computers and is a play on a mythic creature. Harvey is a Púca, a software Púca. Harvey can be fun for those who take time for it. It runs on a little computers. Have an adventure and some fun with Harvey.
Security Note
Harvey is experimental — a working proof of concept, not production-ready software. Letting a probabilistic model direct command execution is an inherently risky attack surface; Harvey mitigates this with safe mode, workspace sandboxing, permission checks, audit logging, and security reviews with each release, but the risk is never zero. Don't use Harvey where the risks might endanger your data, people, or planet. See SECURITY.md for details.
Features
Language Model Support
- Ollama (primary): connect to an Ollama server on this machine or your local network
- hailo-ollama: the Ollama-shaped server for the Raspberry Pi AI HAT+ 2
- Routes: other Ollama servers on your network, via
@mention
- Multi-model dispatch via
@mention and model aliasing; routing feedback shown in spinner
Core Capabilities
- Interactive terminal REPL sandboxed to a workspace directory
- Auto-execute: tagged code blocks in model replies are written to workspace files automatically
HARVEY.md provides a customizable system prompt per workspace- Context utilization hint
[ctx: N%] in spinner when approaching the model's context window
Knowledge, Memory & Sessions
- Three-silo memory: RAG vector stores, session-experience memory with rolling summaries, and a SQLite knowledge base
- Sessions recorded as human-readable Fountain screenplay (
.spmd) files — replay, continue, or mine them for memories
- Model provenance recorded in session headers for audit and replay accuracy
- Pinned context and conversation summarization to manage the context window
File & Code Support
- Code-aware RAG chunking, documentation extraction, and ANSI syntax highlighting (13 languages)
- Automatic code formatting on
write_file: gofmt, clang-format, black, rustfmt, prettier, and built-in Pascal/Oberon/Basic formatters
- PDF text extraction via poppler; image reading via vision-capable model routes
- Remote RAG ingest:
sftp://, scp://, http://, https:// URIs
- File-reference injection: for models that ignore the tools schema, Harvey pre-injects workspace files mentioned in the prompt so they can still work with file content
Extensibility
- SKILL.md skills, bundled skill sets, and multi-step prompt pipelines
- Git integration and per-workspace profile templates
Quick Start
- Install Ollama from https://ollama.com and pull a model, for example
ollama pull qwen2.5-coder:7b
- Install Harvey: run the installer for your platform
- Linux/macOS:
./installer.sh
- Windows: run
installer.ps1 in PowerShell
- Run:
harvey
- Try it:
harvey > /model use
harvey > /read LICENSE
harvey > /help
See Getting Started and Installation for detailed instructions.
Platform Support
Harvey runs on:
- Linux: x86_64, aarch64, armv7l (including Raspberry Pi OS)
- macOS: Intel and Apple Silicon (M1 and above)
- Windows: x86_64
Documentation
Release Notes
-
version: 0.0.19
-
status: active
-
released: 2026-10-10
-
Behaviour changes, see CHANGES.md upgrade notes: llama.cpp models are looked for in llama.cpp's cache ($LLAMA_CACHE, else ~/.cache/llama.cpp) and ~/Models is for llamafiles only, so set llamacpp.models_dir to keep the old directory (harvey DR-0031, proposed); skills.catalog is off by default (set skills: catalog: full to restore it); the command line now beats harvey.yaml for --llamafile-url, --llamafile-dir, -r, --record-file, --resume, --continue, --replay, --replay-output and --replay-continue as well as --ollama and -m, and HARVEY_LLAMAFILE_DIR beats llamafile.models_dir
-
A chat stream that ends before its final chunk is an error (ErrStreamTruncated, exit 74) instead of an empty reply recorded as a success
-
Hailo (AI HAT+ 2) is an engine beside Ollama (harvey DR-0030): a hailo: section in harvey.yaml, card and server detected apart, engine/model names (a bare name on both engines asks, or is exit 2 unattended), the catalog shown as not pulled with /model pull [--yes] NAME, and a capability table so harvey stops sending tools or asking for embeddings where Hailo has none
-
Per-model prompt limits: /model limit, the system prompt and history fitted to the limit, a turn over it refused before sending, and a failed turn retried on smaller prompts that teaches the limit; the model cache is keyed by engine and model
-
/memory mine reads a session too large for the model in parts, records an empty session as empty, and remembers a model that failed on a session so auto-mine does not retry it; a RAG retrieval failure is reported once per session instead of silently
-
harvey completion bash|powershell [-install] for shell completion of verbs, options and help topics
-
A write offer no longer treats a command, end of input or arbitrary text as an answer; piping /exit used to create a file named exit
-
Switching to a llama.cpp model mid-session keeps the log and records the switch, as the Ollama paths already did
Authors
Software Requirements
- Ollama plus Ollama models
Software Suggestions
- Go >= 1.26.4
- CMTools >= 0.0.45
- Pandoc >= 3.9
- GNU Make >= 3.8
Getting Help & License
