local-first · privacy-first · intelligence-driven

Features

Everything you need, locally

No data leaves your machine. Works with llama.cpp, Ollama, vLLM — any OpenAI-compatible backend.

~

Persistent Memory

Keyword-searchable JSONL-backed memory that survives across sessions. The agent remembers your projects, preferences, and context.

>

Mission Tracking

Set objectives that persist across sessions. Agent tracks progress, maintains focus, and updates its own context layer via out_of_the_box.py.

#

Smart Web Search

Generates multi-query searches via the LLM, aggregates and deduplicates results. SearXNG primary with Mojeek fallback.

/

Precision Editing

Surgical line-level tools — insert_lines, patch_file, replace_lines. No full-file rewrites. Resume truncated writes seamlessly.

&

Code Validation

Built-in syntax checking for Python (AST), JavaScript, HTML, and JSON. Catches errors immediately after every write.

*

Self-Improving Context

The LLM autonomously enriches its own context layer. Every conversation makes it smarter about your workflow over time.

@

Session History

Every session auto-saves. Browse, search, and resume past conversations with full context preserved. 50-session auto-cleanup.

%

Built-in Benchmarks

Run SWE-bench, BigCodeBench, GAIA, and LiveCodeBench through the agent's own ReAct loop — no separate tooling needed.

Philosophy

Designed for the terminal

Four principles that guide every design decision.

01

Local First

Your files, code, and memory are primary. Web search is a fallback, not a dependency. Everything runs on your hardware.

02

Precision Over Rewrites

Line-level tools — insert, delete, replace, patch — instead of rewriting entire files. Context is scarce; use it wisely.

03

Context Is Scarce

Read files in 20-50 line chunks. Never bloat context. Smart truncation keeps the conversation focused and efficient.

04

Self-Improving

The LLM feeds insights back into out_of_the_box, enriching its own context across sessions. Every conversation makes it better.

Slash Commands

Control from the prompt

Every interaction is a keystroke away.

/helpList all commands and shortcuts

/historyBrowse and resume past sessions

/codingRun task in structured coding mode

/missionManage persistent objectives

/searchSingle-shot web search

/memorySearch persistent memory

/toolsBrowse tool execution history

/ootbView dynamic context layer

/statusShow endpoint URLs and paths

/resumeContinue most recent session

/sessionLoad a session by ID

/skillsLoad SKILL.md definitions

Keyboard Shortcuts

Ctrl+O Tool history Ctrl+N New conversation Ctrl+/ Shortcuts help Enter Send message Shift+Enter New line ArrowUp Edit last message

Install

Choose your setup

pip, source, or Docker — pick what works for you.

# Install via automated installer (macOS/Linux/WSL) $ curl -fsSL https://raw.githubusercontent.com/oppenheimer-rick/open-agent/main/install.sh | bash # After installation, run the agent using the shortcuts: $ openagent # or simply: $ op

# install from PyPI $ pip install open-agent-cli # configure your LLM backend $ export LLM_BASE="http://localhost:8083/v1" # start the interactive REPL $ openagent

# clone and build $ git clone https://github.com/oppenheimer-rick/open-agent $ cd open-agent # create venv and install $ make venv $ source venv/bin/activate # run the agent $ openagent

# clone and start $ git clone https://github.com/oppenheimer-rick/open-agent $ cd open-agent # start services $ docker compose up -d # run the agent $ docker compose exec openagent openagent

Compare

How it stacks up

Open-Agent vs proprietary cloud CLI agents.

	open-agent	Claude Code	Gemini CLI
Any local LLM	✓	✗	✗
Fully private	✓	✗	✗
Self-improving context	✓	✗	✗
Persistent mission tracking	✓	✗	✗
Session history + search	✓	✓	✗
Built-in code validation	✓	✗	✗
Loop/tool repetition guard	✓	✗	✗
Built-in benchmarks	✓	✗	✗
Precision line editing	✓	✓	✓
Cost	Free	Per-token	Per-token

Get started in seconds

No signup. No API keys. Just a terminal and a local LLM.

Install Now

OpenAgent