bbarit-agent-oss

What bbarit-oss is, in one line. Pi's simplicity + qwen-code's memory + a project wiki + Rust + Pi's gaps closed (auto-loads Claude Code & Codex MCP servers and skills) + agent personas.

한 줄 요약. Pi의 단순함 + qwen-code의 메모리 + 프로젝트 위키 + Rust + Pi 단점 보완(Claude·Codex의 MCP 서버·스킬 자동 로드) + 에이전트 페르소나.

Origin. bbarit-oss began as the agent built into BBARIT Terminal, our desktop AI coding IDE, and is now extracted and released as a standalone open-source CLI.

Why it's built on Pi (MIT). We deliberately inherited Pi's core philosophy — a small, legible agent loop and a tight set of first-class tools — rather than reinventing the loop or reaching for a heavy framework. A coding agent has to be predictable and debuggable, and Pi proved a minimal core can still be capable. We kept that simplicity and its provider-agnostic model registry, then built well beyond it: a multi-process orchestrator for parallel sub-agents, a project wiki, 295 personas, and bundled semantic code search. It is a from-scratch Rust rewrite — no Pi source lines (comment overlap 0.1%, prose 0%); only non-copyrightable design and data (model IDs, endpoints) are shared.

Why the memory design comes from qwen-code (Apache-2.0). Pi has no cross-session memory. Instead of inventing a weaker ontology, we adopted qwen-code's proven auto-memory design — its durable-fact taxonomy (user / feedback / project / reference) and human-readable MEMORY.md index — and wrote the Rust implementation and prompts ourselves.

Why Rust. One self-contained static binary — no Node, no Python, no runtime to install. Startup is instant, memory stays low, and parallel sub-agents plus live streaming are fast and memory-safe. It cross-compiles to macOS, Linux, and Windows as a single file — one-line installers on every platform (curl … | sh on macOS/Linux, irm … | iex on Windows) — and --upgrade self-updates in place everywhere.

Why a built-in wiki. Left to itself, an agent forgets your codebase between sessions and burns time and tokens re-deriving how it works. The project wiki gives it a durable, human-readable markdown knowledge base about the code — architecture, subsystems, gotchas — so it stops re-exploring from scratch and stays consistent turn after turn. (Distinct from auto-memory: the wiki is knowledge about the code; memory is facts about you and the project.)

Why it can reuse Claude Code & Codex MCP servers and skills (closing a Pi gap). Re-registering MCP servers and skills for every new agent is friction Pi never removed. Turn on /interop and bbarit-oss reads your existing ~/.claude.json / ~/.claude/skills and ~/.codex/config.toml / ~/.codex/skills and reuses them as-is — zero reconfiguration. It is off by default so a fresh install stays lean: no foreign servers spawned, no prompt tokens spent on other tools' skill libraries.

Why agent personas. A generic assistant is mediocre at everything. 295 curated personas turn the agent into a domain specialist — a backend architect, a security auditor, a data-viz designer — each a full personality brief (expertise, working style, priorities, taboos), not a one-line "act as X."

The exact measured overlap and every difference are disclosed in PROVENANCE.md; attributions are in NOTICE.

한국어 — 왜 Pi를 택했나. Pi의 핵심 철학, 즉 작고 읽히는 에이전트 루프와 최소한의 1급 도구 세트를 의도적으로 물려받았습니다 — 루프를 재발명하거나 무거운 프레임워크에 기대는 대신. 코딩 에이전트는 예측 가능하고 디버그 가능해야 하며, Pi는 최소 코어로도 충분히 유능할 수 있음을 증명했습니다. 그 단순함과 프로바이더 불가지론 모델 레지스트리는 그대로 유지하되, 그 위에 멀티프로세스 오케스트레이터, 프로젝트 위키, 295 페르소나, 번들 시맨틱 코드검색을 크게 쌓았습니다. Rust로 처음부터 재작성했으며(Pi 소스 0줄 — 주석 겹침 0.1%, 창작 텍스트 0%), 저작권 대상이 아닌 설계·데이터(모델 ID·엔드포인트)만 공유합니다.

한국어 — 왜 qwen-code의 메모리인가. Pi에는 세션 간 메모리가 없습니다. 더 약한 온톨로지를 새로 만들기보다, qwen-code의 검증된 자동 메모리 설계 — 지속 사실 분류 (user / feedback / project / reference)와 사람이 읽는 MEMORY.md 인덱스 — 를 채택하고, Rust 구현과 프롬프트는 직접 작성했습니다. 출처는 NOTICE에 명시.

한국어 — 왜 Rust로 개발했나. 런타임 없이 도는 단일 정적 바이너리 — Node·Python·node_modules 불필요. 시작이 즉각적이고 메모리가 낮으며, 병렬 서브에이전트와 실시간 스트리밍이 빠르고 메모리 안전합니다. macOS·Linux·Windows로 크로스컴파일되는 단일 파일 — 모든 플랫폼 원라인 설치(macOS·Linux는 curl … | sh, Windows는 irm … | iex) — --upgrade는 모든 플랫폼에서 제자리 자기업데이트됩니다.

한국어 — 왜 위키를 쓰나. 그냥 두면 에이전트는 세션이 바뀔 때마다 코드베이스를 잊고 구조를 다시 파악하느라 시간과 토큰을 낭비합니다. 프로젝트 위키는 코드에 대한 지속적·사람이 읽는 마크다운 지식베이스(아키텍처·서브시스템·함정)를 에이전트에 줘서, 매번 처음부터 재탐색하는 걸 멈추고 턴을 넘어 일관성을 유지하게 합니다. (자동 메모리와 구분: 위키는 코드에 대한 지식, 메모리는 당신과 프로젝트에 대한 사실입니다.)

한국어 — 왜 Claude·Codex의 MCP 서버·스킬을 자동 로드하나 (Pi 단점 보완). 새 에이전트마다 MCP 서버와 스킬을 다시 등록하는 건 Pi가 없애지 못한 마찰입니다. bbarit-oss는 기존 ~/.claude.json / ~/.claude/skills, ~/.codex/config.toml / ~/.codex/skills를 그대로 읽어 재사용합니다 — 기본 ON(/interop)이라 재설정 없이 첫날부터 도구가 작동합니다.

한국어 — 왜 에이전트 페르소나인가. 범용 비서는 모든 걸 어중간하게 합니다. 295개 큐레이션 페르소나가 에이전트를 도메인 전문가(백엔드 아키텍트·보안 감사자·데이터 시각화 디자이너)로 바꿉니다 — "act as X" 한 줄이 아니라 전문성·업무 스타일·우선순위· 금기를 담은 완전한 성격 브리프입니다.

bbarit-oss is a fast, terminal-native AI coding agent — an open-source CLI that reads, writes, and edits your code, runs your shell, searches your repository, and pair-programs with the LLM of your choice. It ships as a single static Rust binary with no runtime to install, and works with Anthropic Claude, OpenAI GPT / Codex, Google Gemini, and a dozen more providers (plus local models via Ollama) from one unified model registry.

Think of it as a self-hostable, provider-agnostic alternative to Claude Code, Codex CLI, and Gemini CLI — you own the keys, the data, and the binary.

Table of contents

• Why bbarit-oss?
• Install
• Quick start
• Usage
• Providers & authentication
• Tools
• Slash commands
• Personas
• Auto-memory
• Project wiki
• Sessions
• Skills, extensions, LSP & MCP
• Configuration
• Self-update
• Comparison
• How it works
• Contributing
• FAQ & troubleshooting
• Credits & license

Why bbarit-oss?

• 🦀 One static binary, no runtime. No Node, no Python, no node_modules. Download and run — startup is instant.
• 🔌 Bring your own model. Anthropic, OpenAI (+ Codex), Google (Gemini / Vertex), OpenRouter, Groq, Mistral, Together, Fireworks, DeepSeek, Cerebras, Amazon Bedrock, GitHub Copilot, and local models via Ollama — 1,000+ models from one registry, switchable mid-session.
• 🔒 Private by default. Your keys and code stay on your machine. Nothing is hard-coded, no secrets ship in the binary, and it never phones home. Point it at a local model and stay fully offline.
• 🧠 A real agent, with real tools. It reads and edits files, runs your shell, greps and finds, does semantic code search over your repo, fetches the web, and spawns parallel sub-agents — in an autonomous tool-use loop.
• 🎭 295 built-in personas across 30 domains — turn the agent into a specialist (backend, SRE, security, data, design, …) with one flag.
• 🧩 Extensible & standard. Local skills, extensions, LSP servers, and MCP servers plug straight in.
• 🖥️ A genuinely nice TUI. Word-wrapped transcript, live token streaming, syntax highlighting, model/login pickers, themes, and shell-style history.

Install

One line (macOS / Linux)

curl -fsSL https://bbarit.com/agent/install.sh | sh

Downloads a prebuilt bbarit-oss binary for your platform and installs it into ~/.local/bin (override with BBARIT_INSTALL_DIR).

One line (Windows, PowerShell)

irm https://bbarit.com/agent/install.ps1 | iex

Installs into %LOCALAPPDATA%\Programs\bbarit-oss (override with BBARIT_INSTALL_DIR) and adds it to your user PATH. Binaries are also published on the releases page.

From source

Requires the Rust toolchain (stable).

git clone https://github.com/bbarit/bbarit-agent-oss
cd bbarit-agent-oss
cargo build --release
./target/release/bbarit --help

Supported platforms

Quick sta