Harvey User Manual

Harvey is a tool for scholarly work via natural language programming. It provides an interactive REPL backed by a local language model system — running via llamafile or Ollama — where you direct the model to read files, run commands, search code, and apply changes inside a sandboxed workspace. Individual prompts can be routed to cloud providers via named routes. Language model systems are commonly called “AI models” or “AI”; Harvey treats them as a programmable interface for deliberate, documented work rather than a chat assistant.

🚀 Quick Start Guide

For users who want to get Harvey running immediately

1. Installation

INSTALL.md — Step-by-step instructions to install Harvey on Linux, macOS, or Windows

2. First Session

getting_started.md — Launch Harvey, connect to a local model, and run your first commands

3. System Prompt Customization

HARVEY.md — The default system prompt and how to customize it for your project

📚 Core User Guide

For daily Harvey usage

Working with Harvey

• Overview — What Harvey is, its philosophy, and core features
• Configuration — Workspace configuration in agents/harvey.yaml, including permissions, timeouts, and memory settings

Commands and Workflows

• Command Reference — Primary man page with all available commands and flags
• Getting Started — Detailed session walkthrough, keyboard shortcuts, and slash commands

Command vocabulary

Every Harvey resource command responds to the same eight verbs:

Backend service commands additionally support start, stop, and status (connection health — distinct from show which displays content).

The add vs new distinction: add registers something you already have; new creates something Harvey manages from scratch.

Learning this vocabulary for one command teaches you all the others: /llamafile, /rag, /route, /session, /skill, /skill-set, and /memory profile all follow the same pattern.

🎯 Advanced Features

Retrieval-Augmented Generation

• RAG Setup & Usage — Complete guide to retrieval-augmented generation, including named stores, ingest commands, and configuration
• RAG Commands — Detailed reference for all RAG-related slash commands

Knowledge Management

• Knowledge Base — SQLite-backed knowledge base schema, FTS5 search, CLI commands, and Go API
• KB Commands — /kb slash command reference: projects, observations, concepts
• Memory System — Unified memory architecture with rolling summaries and token budget tracking
• Memory Commands — /memory slash command reference: mine, list, show, flag, forget, profile
• Learn — Memory Overview — When to use RAG vs memory vs KB; unified recall

Model Management

• Llamafile Commands — Primary local backend: add, use, show, list, start, download, remove
• Unified /model Command — /model list|use|show|status works across llamafile and Ollama
• Model & Alias Commands — @mention inline model switching and /model alias short names
• Ollama Commands — Alternative local backend: service control and model management
• Routing — Connect to remote endpoints (Anthropic, DeepSeek, Gemini, Mistral, OpenAI, remote Ollama) via @mention syntax
• Routing Commands — /route slash command reference
• Model Guide — Model selection guide based on capability probing results
• Model Cache — Model capability caching architecture, database schema, and probing mechanisms

Sessions & Recording

• Sessions — Session recording, file structure, replay functionality, and programmatic access
• Session Commands — /session list|show|use|continue|replay reference
• Record Commands — /record start|stop|status reference
• Fountain Format — Fountain screenplay format specification for conversation recordings

File & Workspace Operations

• Read — Inject workspace files into conversation context
• Read Directory — Read all files in a directory tree into context
• Read PDF — Extract and inject PDF text (requires poppler)
• Attach — Attach images, PDFs, or text; auto-selects best representation
• Write — Save the last reply or first code block to a file
• Format — Auto-format source files using language-appropriate tools
• Search — Regex search across workspace files
• Files — List workspace directory contents
• File Tree — Display a recursive directory tree

Automation & Pipelines

• Pipeline — Chain Markdown prompt files with confidence gating
• Plan — Generate and execute step-by-step task plans with bounded context
• Loop — Repeat a prompt or command on a fixed interval

Status & Diagnostics

• Status — Show backend, token usage, routing, recording, and debug state
• Hint — Actionable suggestions for improving results (RAG, memory, KB)
• Inspect — Detailed model information (Ollama)
• Summarize — Condense history to free context window space (/compact alias)

🔧 Extending Harvey

Skills System

• Skills Overview — Extend Harvey with custom skills, including SKILL.md format, discovery paths, compiled skills, triggers, and the skill wizard
• Skills Commands — /skill slash command reference
• Skill Sets — Manage named bundles of skills for specific workflows

Security

• Security Guide — Safe mode, command allowlists, workspace permissions, audit logging, and threat model
• Security Commands — /safemode, /permissions, /audit, /security reference
• Security Review — Pre-release security assessment and known limitations

📖 What’s New

• See CHANGES.md for version history and new features in each release

🔍 Can’t Find What You Need?

• Use /help inside Harvey for a live command list
• Use /help <command> for specific command help
• See Developer Guide for development and contribution information
• See Reference for complete command and configuration reference