Harvey

Harvey is an agent REPL written in Go and designed to use Llamafile models or 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

• Llamafile (primary): register and run local model binaries — no server required
• Ollama: connect to a local Ollama server for broader model selection
• Cloud routes: Anthropic, DeepSeek, Gemini, Mistral, and OpenAI via configured routes
• 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: s3://, 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

1. Download a llamafile from the Mozilla AI pre-built models page and make it executable

2. Install Harvey: run the installer for your platform

   • Linux/macOS: ./installer.sh
   • Windows: run installer.ps1 in PowerShell

3. Run: harvey

4. 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

• User Manual — Main documentation index
• Overview — What Harvey is, why you might use it, and the design philosophy
• Getting Started — First session, keyboard shortcuts, slash commands
• Configuration Reference — harvey.yaml fields and environment variables
• Installation — Get Harvey installed on your system
• Developer Guide — Architecture, conventions, and contributing
• Vision — Philosophy, motivation, and future direction
• About — Project metadata and version information
• GitHub Repository — Source code and issues

Release Notes

• version: 0.0.17

• status: active

• released: 2026-09-27

• Breaking: harvey and assay now use the workspace exit-code convention (workspace DR-0003; harvey DR-0005/DR-0006) instead of exiting only 0 or 1. Five cases that exited 0 now fail (--continue/--record-file naming a bad path, a malformed harvey.yaml or no reachable backend in a non-interactive session, and assay when every model call fails); every other case exits a more specific code than a bare 1. See harvey/CHANGES.md’s v0.0.17 section for the full upgrade table

• -json/--json on both binaries: a failing startup/flag-parsing error (or, for harvey, a failed non-interactive session) prints as {"error","class","code"} JSON on stderr instead of text, matching kb’s shape

• A non-interactive harvey session (piped stdin, or --replay without --replay-continue) now exits with the class of its first failed slash command or chat turn, instead of always 0; an interactive session at a terminal is unaffected. /plan’s errors are the first command family reclassified to give real signal here

• /rag setup’s embedding-model auto-pick now confirms its keyword-based guess with one live /api/embed call before committing to it, falling back to another candidate (or warning) when the guess is wrong — ThoroughProbeModel’s first caller

• Ollama models are now probed on every selection, not just when a new alias is saved, so /model use NAME and the startup picker no longer silently degrade to file-injection instead of tool calls; a /model mode override survives re-probing (previously erased)

• Removed roughly 30 stale /ollama//llamafile references across runtime messages, help text and generated docs; both commands were replaced by /model months ago. stale_commands_test.go guards against regressions

• Bug fix: closed stdin (Ctrl-D, a closed pipe) at a yes/no prompt no longer defaults to “yes” — affected Start Ollama now?, Restart MODEL?, and three skill prompts

• Bug fix: a failed memory-embed no longer leaves an orphaned .fountain file; the error now says to run ollama pull MODEL

• Bug fix: /model use NAME now matches any registered model by exact name or unique prefix, not just the picker’s list

• Bug fix: the “1-byte” line reader actually buffered 16 bytes (bufio’s minimum) and could swallow piped multi-line input after a prompt

• knowledge bumped to the released v0.0.14; termlib bumped to the released v0.0.10 (a display-corrupting bug fix for prompts wider than the terminal or containing an embedded newline)

Authors

• Doiel, R. S.

Software Requirements

• Llamafile v0.10 models or Ollama plus Ollama models

Software Suggestions

• Go >= 1.26.4
• CMTools >= 0.0.45
• Pandoc >= 3.9
• GNU Make >= 3.8

Getting Help & License

• GitHub Issues — Bug reports and feature requests
• AGPL-3.0 License