Production-grade LLM Agent Kernel
Multi-mode orchestration · Mathematical context compression · Built-in safety · External Agent System
v2.7.0 · MIT License · Rust 1.75+ · Bun + OpenTUI · macOS / Linux
Abacus is a terminal-native LLM agent kernel that orchestrates AI reasoning across four collaborative modes. It provides a rich TUI experience with real-time streaming, tool execution, and multi-expert consultation — all from your terminal.
v2.7.0 New Features:
- TypeScript TUI — Migrated from Rust ratatui to OpenTUI (Zig-native rendering) + XState + Zustand
- External Agent System — Install and manage external agents via MCP protocol
- OKLCH Theme System — 11 themes with perceptually uniform colors
- Scene-Aware UI — Trading/Coding/Editing scenarios with adaptive layouts
- 21 Tool Clusters — Full coverage of 72 tools with intelligent routing
Clarify Mode — single-agent deep reasoning:
┌─────────────────────────────────────────────────────────────────────────────────┐
│ ⠋ ABACUS ▸ Refactor auth module · 澄清 deepseek-v4 │
├──────────────────────────────────────────────────────────┬──────────────────────┤
│ │ 📊 [====------] 38% │
│ User: 帮我把 session 认证改成 JWT │ 415K↑ 5K↓ c78% │
│ │ 🔧 12/14 · 调8✓7 │
│ Session: │ 🧠 3域 12行为 │
│ ## 方案确认 │ ╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌ │
│ 我决定使用 JWT token 替代 session cookie。 │ 现场 │
│ 原因：无状态、易扩展、前后端解耦。 │ ▸ 12:03 分析代码 │
│ │ ⚙ fs_read auth.rs │
│ ```rust │ ▸ 12:04 方案生成 │
│ impl AuthService { │ ✓ 2 工具完成 │
│ pub fn verify_token(&self, token: &str) │ ╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌ │
│ -> Result<Claims, AuthError> │ Focus · 澄清 · 3轮 │
│ } │ JWT 认证重构 │
│ ``` │ → 方案已确认 │
│ │ │
├──────────────────────────────────────────────────────────┬──────────────────────┤
│ ╭──────────────────────────────────────────────────────╮ │ deepseek(✓) · v4 │
│ │ ⠋ Thinking 澄清 · 分析认证模块 3.2s │ │ ⬡ 52K/128K/1M · 3轮 │
│ │ Ask anything... ⏎ Enter │ │ thinking · ¥0.08 │
│ ╰──────────────────────────────────────────────────────╯ │ │
├──────────────────────────────────────────────────────────┴──────────────────────┤
│ ● 澄清 1.2K tok Cmd+↑↓ Ctrl+B Esc │
└─────────────────────────────────────────────────────────────────────────────────┘
Plan Mode — two-phase: research → approval → execute:
┌─────────────────────────────────────────────────────────────────────────────────┐
│ ● ABACUS ▸ Refactor auth module · 澄清 · [📋 规划] deepseek-v4 │
├──────────────────────────────────────────────────────────┬──────────────────────┤
│ │ 📊 [======----] 58% │
│ Session: │ 820K↑ 12K↓ c65% │
│ 📋 计划已就绪 — 3 阶段, 12 步骤 │ 🔧 14/14 · 调12✓10 │
│ │ ╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌ │
│ 选择执行策略: │ Focus · 规划 2/3 │
│ [A] 自动执行 — 工具调用自动放行 │ 目标 重构auth模块 │
│ [S] 逐步确认 — 每步操作需确认 │ ✓ 分析现有代码 │
│ [T] 团队分发 — 多专家并行执行 │ ⟳ 生成迁移方案 │
│ [C] 取消 │ ○ 执行重构 │
│ │ ████░░░░░░ 2/3 │
│ 输入 A/S/T/C: │ │
│ │ │
├──────────────────────────────────────────────────────────┬──────────────────────┤
│ ╭──────────────────────────────────────────────────────╮ │ deepseek(✓) · v4 │
│ │ ● Ready · 澄清 │ │ ⬡ 75K/128K/1M · 8轮 │
│ │ 输入 A/S/T/C 选择策略... ⏎ Enter │ │ thinking · ¥0.32 │
│ ╰──────────────────────────────────────────────────────╯ │ │
├──────────────────────────────────────────────────────────┴──────────────────────┤
│ ● 澄清 · 📋 规划就绪 4.8K tok Cmd+↑↓ Ctrl+B Esc │
└─────────────────────────────────────────────────────────────────────────────────┘
Meeting Mode — multi-expert consultation:
┌─────────────────────────────────────────────────────────────────────────────────┐
│ ⠙ ABACUS ▸ 性能优化方案 · 会诊 deepseek-v4 │
├──────────────────────────────────────────────────────────┬──────────────────────┤
│ │ 📊 [=======---] 72% │
│ User: @架构师 @DBA 这个查询太慢了怎么优化 │ 1.1M↑ 8K↓ c80% │
│ │ 🔧 14/14 · 调15✓12 │
│ 🔊 架构师: │ ╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌ │
│ 建议从两个层面入手： │ Focus · 会诊 2/3位 │
│ 1. 应用层：添加 Redis 缓存热点查询 │ 🔊 架构师 (后端) │
│ 2. 数据层：对 user_id 添加复合索引 │ 🔊 DBA (数据库) │
│ │ ○ 前端 (前端) │
│ 🔊 DBA: │ 阶段 ● 发言中 │
│ 补充几点： │ │
│ - EXPLAIN 显示全表扫描，缺少 (user_id, created_at) 索引 │ │
│ - 建议分区表（按月），历史数据走冷存储 │ │
│ │ │
├──────────────────────────────────────────────────────────┬──────────────────────┤
│ ╭──────────────────────────────────────────────────────╮ │ deepseek(✓) · v4 │
│ │ ⠹ Working 会诊 · DBA 发言中 5.1s │ │ ⬡ 92K/128K/1M · 5轮 │
│ │ Ask anything... ⏎ Enter │ │ thinking · ¥0.56 │
│ ╰──────────────────────────────────────────────────────╯ │ │
├──────────────────────────────────────────────────────────┴──────────────────────┤
│ ● 会诊 · DBA 发言中 6.2K tok Cmd+↑↓ Ctrl+B Esc │
└─────────────────────────────────────────────────────────────────────────────────┘
Abacus is a pre-compiled binary with no runtime dependencies. You only need:
- macOS 12+ (Monterey) or Linux (glibc 2.31+, e.g. Ubuntu 20.04+)
- A terminal emulator (Terminal.app, iTerm2, Warp, Kitty, Alacritty, etc.)
- An LLM API key (DeepSeek / OpenAI / Anthropic / any OpenAI-compatible endpoint)
For building from source, you additionally need:
- Rust 1.75+ (
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh) - A C linker (macOS: Xcode Command Line Tools; Linux:
build-essential) - Protocol Buffers compiler (
brew install protobufon macOS;apt install protobuf-compileron Linux)
curl -fsSL https://github.com/lucasli000/abacus/releases/latest/download/install.sh | shThis auto-detects your platform (macOS/Linux, ARM64/x86_64), downloads the correct binary, and installs to /usr/local/bin/.
To install to a custom path:
INSTALL_DIR=~/.local/bin curl -fsSL https://github.com/lucasli000/abacus/releases/latest/download/install.sh | shGo to Releases and download for your platform:
| Platform | File | Size |
|---|---|---|
| macOS Apple Silicon (M1/M2/M3/M4) | abacus-aarch64-apple-darwin.tar.gz |
~19 MB |
| Linux x86_64 | abacus-x86_64-unknown-linux-gnu.tar.gz |
~22 MB |
Then install:
# 1. Extract
tar -xzf abacus-aarch64-apple-darwin.tar.gz
# 2. Move to PATH
sudo mv abacus /usr/local/bin/
# 3. (macOS only) Remove quarantine flag
xattr -d com.apple.quarantine /usr/local/bin/abacus 2>/dev/null || true
# 4. Verify
abacus# Install Rust (if not already)
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
source ~/.cargo/env
# Clone and build
git clone https://github.com/lucasli000/abacus.git
cd abacus/pkg
cargo build --release --package abacus-cli
# Install
sudo cp target/release/abacus /usr/local/bin/# Start Abacus — it launches a full-screen TUI
abacusOn first run, a setup wizard will guide you through:
- Choosing an LLM provider (DeepSeek / OpenAI / Anthropic / custom)
- Entering your API key
- Selecting a default model
Config is saved to ~/.abacus/config.yaml. You can skip the wizard in CI:
export ABACUS_API_KEY=sk-xxxIf Abacus launches and shows the TUI interface, installation is successful. The program uses the full terminal screen — press Ctrl+D or type /quit to exit.
rm /usr/local/bin/abacus
rm -rf ~/.abacus # Remove config + data (optional)| Problem | Solution |
|---|---|
zsh: killed abacus |
macOS code signing issue. Run: codesign --sign - --force /usr/local/bin/abacus |
permission denied |
Run: chmod +x /usr/local/bin/abacus |
command not found |
Ensure /usr/local/bin is in your $PATH |
"abacus" cannot be opened (macOS) |
Run: xattr -d com.apple.quarantine /usr/local/bin/abacus |
Error in the HTTP2 framing layer |
Network issue (common in China). Use mirror: see below |
| Blank screen / no output | Ensure your terminal supports alternate screen (most modern terminals do) |
China mainland users — if downloads fail due to network issues:
# Option A: Use mirror proxy
curl -fsSL https://gh-proxy.com/https://github.com/lucasli000/abacus/releases/download/v2.5.5/abacus-aarch64-apple-darwin.tar.gz | tar -xz
sudo mv abacus /usr/local/bin/
# Option B: Use local SOCKS proxy
curl -x socks5://127.0.0.1:7890 -fsSL https://github.com/lucasli000/abacus/releases/latest/download/install.sh | sh| Mode | Description | Use Case |
|---|---|---|
| Clarify | Single-agent deep reasoning with progressive gate | Default — ask questions, write code, debug |
| Plan | Two-phase: Research → Approval → Execute | Complex multi-step tasks, user selects strategy (A/S/T) |
| Team | Multi-agent parallel with ToolAgent delegation | Large refactoring, parallel tool operations |
| Meeting | Multi-expert weighted consultation | Cross-domain problems needing diverse expertise |
- Mathematical Context Compression — Information Bottleneck + H2O Heavy-Hitter + ARC + Greedy Knapsack. 38% message reduction, 100% key decision retention.
- ToolActionClassifier — Rule-based safety (hard_deny/soft_deny/allow). Zero LLM overhead.
- ToolAgent — Batch delegation for read-only ops. 4 built-in types: Explorer, Researcher, Coder, Mathematician.
- MCIP — Multi-level permission gate (role → confirm → capability).
- ProgressiveGate — Complexity-aware output strategy.
- Dynamic Timeout — Complexity + tool count + LLM self-extension.
- Self-Learning — EffectivenessTracker records tool outcomes. Experience injector marks unreliable tools. Anti-pattern detector catches environment issues.
- Unicode Injection Defense — Zero-width character normalization prevents prompt injection bypass.
- Configurable Safety Rules — User-defined injection patterns with priority override.
- Real-time streaming with thinking visualization
- Soft-wrap input with cursor tracking
- Right panel: dashboard + timeline + focus
- 12 built-in themes
- Inline suggestions (Tab)
- Honest Data Display — No fake data when provider/model not connected. Shows "Not connected" instead of placeholders.
- Formatted Tool Output — JSON parsed into structured display:
⚙ command+│ ⤷ stdout - Symbol System — Consistent visual language: ▸ You / ◈ Think / ⚙ Tool / ◇ Abacus
- Markdown in Thinking — Thinking cards render markdown with syntax highlighting
- Right-aligned Metadata — Model name and timestamps aligned to card header right side
# Interactive TUI (default)
abacus
# Single query
abacus ask "explain this error"
# Specific model
abacus --model deepseek-v4 ask "optimize this"
# Mode switching inside TUI
/clarify # Single-agent (default)
/plan <goal> # Two-phase planning
/team <task> # Multi-agent parallel
/meeting # Multi-expert consultation
@expert msg # Auto-route to Meeting┌─ Abacus Setup ──────────────────────────────┐
│ │
│ 1. Choose LLM provider │
│ → OpenAI / Anthropic / DeepSeek / ... │
│ │
│ 2. Enter API key │
│ │
│ 3. Select default model │
│ │
└──────────────────────────────────────────────┘
Auto-detects 9 providers from URL. Config: ~/.abacus/config.yaml
abacus/pkg/
├── crates/
│ ├── abacus-cli/ TUI + CLI (ratatui + clap)
│ ├── abacus-core/ Engine kernel (pipeline, context, tools, safety)
│ ├── abacus-orchestrator/ Meeting/Team orchestration
│ ├── abacus-types/ Shared types
│ └── abacus-server/ HTTP/SSE server (axum)
├── assets/ Logo + icons
└── scripts/ Install script
| Provider | Models | Status |
|---|---|---|
| OpenAI | GPT-4o, o1, o3 | ✅ |
| Anthropic | Claude Sonnet/Opus | ✅ |
| DeepSeek | V3, V4, R1 | ✅ |
| Gemini 2.x | ✅ | |
| Moonshot / 智谱 / 通义 / SiliconFlow / Groq | Various | ✅ |
| Any OpenAI-compatible | Custom | ✅ |
cargo check --workspace # Type check
cargo test --workspace # Run tests (1814 tests)
cargo build --release --package abacus-cli # Release build- 1814 unit tests across all crates
- 0 warnings, 0 errors
- Theme WCAG contrast validation (12 themes × 4 color roles)
MIT — see LICENSE