🪶 Quill — Open-Source AI Super-Agent Framework

The open-source alternative to OpenAI Codex, Cursor, Claude Code, and DeerFlow. AI agent framework with sandboxed code execution, sub-agent orchestration, MCP, skills marketplace, session forking, FTS5 search, multi-agent teams, and a native Tauri desktop app.

English · 中文 · 한국어 · 日本語 · Français · Русский · Español · العربية

](https://opensource.org/licenses/Apache-2.0)

Website · Docs · Quick Start · Desktop App · Skills · Contributing

---

🤔 Why Quill?

Quill is a super-agent framework — an AI that can research, code, analyze data, generate documents, and orchestrate sub-agents to do almost anything. Unlike closed-source alternatives, Quill is fully open-source, self-hostable, and extensible.

Feature	Quill	OpenAI Codex	Cursor	Claude Code	DeerFlow	OpenClaw
Open source	✅ Apache 2.0	❌	❌	❌	✅ MIT	✅ MIT
Self-hosted	✅	❌	❌	❌	✅	✅
Sandboxed execution	✅	✅	✅	✅	✅	✅
Sub-agent orchestration	✅	❌	❌	✅	✅	❌
Desktop app	✅ Tauri	✅	✅	❌	❌	❌
MCP integration	✅ Dual-role	✅	✅	✅	✅	❌
Skills marketplace	✅	❌	❌	❌	❌	✅
Session forking	✅	❌	❌	❌	❌	❌
FTS5 search	✅	❌	❌	❌	❌	❌
Multi-agent teams	✅	❌	❌	❌	❌	❌
Workflow engine	✅	❌	❌	❌	✅	❌
Self-improving skills	✅	❌	❌	❌	❌	❌
Adaptive permissions	✅	❌	❌	❌	❌	❌
IM channels (5+)	✅	❌	❌	❌	✅	✅
Tool receipts	✅	❌	❌	❌	✅	❌
Off-peak deferred tasks	✅	❌	❌	❌	❌	❌
Message queue while busy	✅	❌	❌	❌	❌	❌
Workspace checkpoints / rewind	✅	❌	❌	❌	❌	❌
Read-only command auto-approve	✅	❌	❌	✅	❌	❌
Risk-tiered tool review	✅	❌	❌	❌	✅	❌
Durable kanban board	✅	❌	❌	❌	❌	❌

---

✨ Core Capabilities

🔬 Deep Research

Multi-source search with cross-validation and cited reports. Quill orchestrates multiple sub-agents to investigate topics in depth.

💻 Sandboxed Code Execution

Safely run Python / Bash / file operations in an isolated sandbox environment with full filesystem access. OS-level security with network isolation.

🤖 Sub-Agent Orchestration

Main agent dispatches specialized sub-agents for parallel complex tasks — general-purpose, bash, research, and custom agents. Up to 128 concurrent sub-agents with AgentSwarm.

🧩 Skills Marketplace

Install skills to extend capabilities. Build custom extensions with lifecycle hooks (pre_model, post_model, pre_tool, post_tool). Install from GitHub repos or any URL.

🔀 Session Forking

Branch any conversation at any point. Forked threads carry full conversation history via checkpoint copy — experiment without disrupting the original.

🔍 FTS5 Session Search

BM25-ranked full-text search across all thread messages with snippet extraction and prefix matching. Find any past conversation instantly.

🧠 Long-Term Memory & Dreaming

Continuously records user profile and conversation history with confidence-based fact eviction. Background "dreaming" consolidation promotes short-term signals to durable long-term memory.

🤝 Multi-Agent Teams

Supervisor, round-robin, handoff, and hierarchical team patterns. Shared task board with DAG dependencies and peer messaging.

🔄 Workflow Engine

DAG-based agent orchestration with parallel execution, retry, and conditional branching — plus dynamic workflows: author orchestration as plain TypeScript scripts (top-level await, agent()/ask()/run() facade) with journal-backed amend/resume so finished steps are never re-paid.

🛡️ Safety & Guardrails

Tools declare safety properties (read-only, destructive, idempotent, open-world) that feed into risk-level-based authorization. Deterministic security scanner blocks malicious skills offline. Delegated subagent results are verified by deterministic acceptance criteria (file existence, recorded test runs) — not self-reported.

💓 Proactive Heartbeat

Periodic agent turns that check whether anything needs attention — with a persistent monitor checklist, active-hours windows, and cost-controlled isolated sessions. Scheduled-task outcomes can be delivered straight to your IM (Slack / Feishu / DingTalk / Telegram webhooks).

🌐 Multi-Model & Multi-Platform

DeepSeek / OpenAI / Anthropic / vLLM / Ollama and more, with role→model overlay routing. UI supports 8 languages. IM channels: Telegram, Slack, Discord, Feishu, DingTalk. Conversations can be shared via sanitized, integrity-hashed read-only links — and imported from Claude Code.

🖥️ Native Desktop App

Tauri 2 desktop app with native filesystem access, system tray, workspace sync, and auto-updates. Available for macOS, Windows, and Linux.

---

🚀 Quick Start

Option 1: Desktop App (Recommended)

Download the latest release for your platform from the Releases page.

# macOS
brew install --cask quill  # coming soon

# Or download .dmg/.msi/.AppImage from Releases

Option 2: Local Development

# Clone the repository
git clone https://github.com/CreamyLong/quill.git
cd quill

# Interactive setup wizard (2 minutes)
make setup

# Start all services with hot-reload
make dev

# Open http://localhost:2126 in your browser

Option 3: Docker

docker compose up -d
# Open http://localhost:2126

Option 4: Desktop Development

# One-command desktop (builds frontend, starts Gateway, launches Tauri)
make desktop

# Or manually:
cd desktop
npm install
npm run tauri dev    # first build ~3-5 min, then incremental

# Production build → .dmg/.msi/.AppImage
npm run tauri build

---

📸 Feature Showcase

🔬 Deep Research

Multi-source search with cross-validation and cited reports. Quill orchestrates multiple sub-agents to investigate topics in depth.

💻 Code Execution

Safely run Python / Bash / file operations in an isolated sandbox environment with full filesystem access.

🤖 Sub-Agent Collaboration

Main agent dispatches specialized sub-agents for parallel complex tasks — general-purpose, bash, and custom agents.

🧩 Extensible Skills & Extensions

Install skills to extend capabilities. Build custom extensions with lifecycle hooks (pre_model, post_model, pre_tool, post_tool).

🧠 Long-Term Memory & Dreaming

Continuously records user profile and conversation history with confidence-based fact eviction policies. Background "dreaming" consolidation promotes strong short-term signals to durable long-term memory (Light → REM → Deep phases).

💓 Proactive Heartbeat

Periodic agent turns that check whether anything needs attention — with a persistent monitor scratch checklist, active-hours windows, and cost-controlled isolated sessions.

🛡️ Tool Annotations & Guardrails

Tools declare safety properties (read-only, destructive, idempotent, open-world) that feed into risk-level-based authorization. Deterministic security scanner blocks malicious skills offline before any LLM call.

🌐 Multi-Model & Multi-Language

DeepSeek / OpenAI / Anthropic / vLLM / Ollama and more. UI supports English, 中文, and 한국어.

---

🏗️ System Architecture

High-Level Architecture

graph TB
    subgraph "Client Layer"
        WEB[Next.js Frontend<br/>React + Tailwind]
        IM[IM Channels<br/>Telegram, Slack, Discord<br/>Feishu, DingTalk]
        DESK[Desktop App<br/>Tauri 2]
    end

    subgraph "Gateway Layer (Port 8001)"
        GW[Gateway API<br/>LangGraph Runtime]
        SB[Stream Bridge<br/>SSE Delivery]
        RM[Run Manager<br/>Task Lifecycle]
    end

    subgraph "Agent Runtime"
        LA[Lead Agent<br/>StateGraph]
        MW[Middleware Chain<br/>38+ Middlewares]
        SA[Sub-Agent Executor<br/>Thread Pool]
    end

    subgraph "Infrastructure"
        DB[(Database<br/>SQLite / Postgres)]
        SK[Skills System<br/>SKILL.md + Marketplace]
        MCP[MCP Servers<br/>Dual-Role]
        MEM[Memory System<br/>LLM Extraction + Eviction]
    end

    WEB -->|HTTP/SSE| GW
    IM -->|Webhook| GW
    DESK -->|HTTP/SSE| GW
    GW --> LA
    GW --> SB
    GW --> RM
    LA --> MW
    LA --> SA
    LA -->|Tool Calls| MCP
    LA -->|Read/Write| DB
    LA -->|Load/Save| SK
    LA -->|Extract/Inject| MEM
    SA -->|Background| LA

Loading

Agent Loop & Middleware Chain

flowchart LR
    START([START]) --> PREP[Prepare<br/>Inject System Prompt]
    PREP --> BM[beforeModel<br/>38+ Hooks]
    BM --> MODEL[Model Call<br/>LLM Inference]
    MODEL --> AM[afterModel<br/>Post-Processing]
    AM --> TOOLS{Tool Calls?}
    TOOLS -->|Yes| EXEC[Execute Tools<br/>Sandbox + MCP]
    EXEC --> AA[afterAgent<br/>State Updates]
    AA -->|Continue| BM
    TOOLS -->|No| END([END])
    AA -->|Finish| END

    style START fill:#4ade80,stroke:#166534
    style END fill:#f87171,stroke:#991b1b
    style MODEL fill:#60a5fa,stroke:#1e40af
    style EXEC fill:#fbbf24,stroke:#92400e

Loading

Middleware Pipeline (Lead Agent)

flowchart TB
    subgraph "Input Layer"
        M1[1. Input Sanitization<br/>Prompt Injection Defense]
        M2[2. Tool Output Budget<br/>Size Caps]
        M3[3. Thread Data<br/>Per-Thread Directories]
    end

    subgraph "Context Layer"
        M4[4. Dynamic Context<br/>Date + Memory Reminders]
        M5[5. Skill Activation<br/>/skill-name Slash Commands]
        M6[6. Durable Context<br/>Delegation + Skill References]
    end

    subgraph "Safety Layer"
        M7[7. Guardrail<br/>Pre-Tool Authorization]
        M8[8. Sandbox Audit<br/>Security Logging]
        M9[9. Tool Error Handling<br/>Graceful Recovery]
    end

    subgraph "Model Layer"
        M10[10. LLM Error Handling<br/>Retry + Backoff]
        M11[11. System Message Coalescing<br/>Provider Compatibility]
        M12[12. Deferred Tool Filter<br/>MCP Schema Hiding]
    end

    subgraph "Output Layer"
        M13[13. Summarization<br/>Context Reduction]
        M14[14. Loop Detection<br/>Repetition Breaker]
        M15[15. Token Budget<br/>Per-Run Limits]
        M16[16. Clarification<br/>User Interaction]
    end

    M1 --> M2 --> M3 --> M4 --> M5 --> M6
    M6 --> M7 --> M8 --> M9 --> M10 --> M11 --> M12
    M12 --> M13 --> M14 --> M15 --> M16

    style M1 fill:#dbeafe,stroke:#1e40af
    style M4 fill:#fef3c7,stroke:#92400e
    style M7 fill:#fee2e2,stroke:#991b1b
    style M10 fill:#dcfce7,stroke:#166534
    style M13 fill:#f3e8ff,stroke:#6b21a8

Loading

Sub-Agent Delegation Flow

sequenceDiagram
    participant U as User
    participant LA as Lead Agent
    participant EX as Sub-Agent Executor
    participant SA as Sub-Agent
    participant T as Tools

    U->>LA: Send message
    LA->>LA: Model generates tool_call
    LA->>EX: task(description, type)
    EX->>SA: Create sub-agent graph
    loop Execute turns
        SA->>T: Tool calls (bash, read, write)
        T-->>SA: Results
        SA->>SA: Model reasoning
    end
    SA-->>EX: Final result
    EX-->>LA: task_completed event
    LA-->>U: Response with results

Loading

---

📦 Skills & Extensions Ecosystem

Quill ships with 20+ built-in skills: academic review, deep research, data analysis, PPT generation, chart visualization, image / video / music generation, frontend design, GitHub research, newsletter, and more.

Extensions enable third-party developers to build plugins that hook into the agent lifecycle:

• pre_model / post_model — intercept and modify model calls
• pre_tool / post_tool — intercept and modify tool execution
• on_agent_start / on_agent_end — setup and cleanup

Connect additional MCP services via extensions_config.json.

Install from Marketplace

# Install from GitHub repo
curl -X POST http://localhost:8001/skills/marketplace/install \
  -H "Content-Type: application/json" \
  -d '{"source": "github", "target": "owner/repo"}'

# Install from URL
curl -X POST http://localhost:8001/skills/marketplace/install \
  -H "Content-Type: application/json" \
  -d '{"source": "url", "target": "https://example.com/skill.md"}'

---

🛠️ Tech Stack

Layer	Technology
Frontend	Next.js 15 · React 19 · Tailwind CSS · shadcn/ui
Backend	LangGraph · TypeScript · node:sqlite
Desktop	Tauri 2 (Rust + WebView)
Database	SQLite / PostgreSQL · LangGraph Checkpointer
Agent Runtime	StateGraph · 38+ Middlewares · Sub-Agent Executor
Protocols	MCP (Model Context Protocol) · SSE · HTTP/SSE/Stdio

---

🌐 Internationalization

Quill supports eight languages:

Language	Locale	Status
English	en-US	✅ Complete
中文 (Chinese)	zh-CN	✅ Complete
한국어 (Korean)	ko-KR	✅ Complete
日本語 (Japanese)	ja-JP	✅ Complete
Français (French)	fr-FR	✅ Complete
Русский (Russian)	ru-RU	✅ Complete
Español (Spanish)	es-ES	✅ Complete
العربية (Arabic)	ar-SA	✅ Complete

Switch languages in Settings → Appearance → Language.

---

📊 Competitive Analysis

Quill was systematically evaluated against 10 leading harness frameworks. See harness-framework-comparison.md for the full analysis.

Quill matches or exceeds all 10 frameworks on 18 of 22 capability dimensions.

Framework	Stars	LangGraph	Sandbox	Memory	MCP	Sub-Agent	Desktop	Teams	Search
Quill	⭐	✅	✅	✅	✅	✅	✅	✅	✅
DeerFlow	⭐⭐⭐⭐⭐	✅	✅	✅	✅	✅	❌	❌	❌
OpenWork	⭐⭐⭐	❌	❌	❌	✅	❌	✅	❌	❌
DeepSeek Harness	⭐⭐⭐	❌	✅	✅	❌	✅	❌	✅	❌
Kimi Code	⭐⭐⭐⭐	❌	✅	✅	✅	✅	❌	❌	❌
OpenAI Codex	⭐⭐⭐⭐⭐	❌	✅	❌	✅	❌	✅	❌	❌
CrewAI	⭐⭐⭐⭐⭐	❌	❌	✅	✅	✅	❌	✅	❌
AutoGen	⭐⭐⭐⭐	❌	✅	❌	✅	✅	❌	✅	❌
OpenClaw	⭐⭐⭐	❌	✅	❌	❌	❌	❌	❌	❌
Hermes Agent	⭐⭐⭐	❌	✅	✅	✅	✅	❌	❌	✅

---

🤝 Contributing

Issues and PRs are welcome! See CONTRIBUTING.md for details.

Areas where we especially need help:

• 🌍 Translations — help us support more languages
• 🧩 Skills — create and share new skills
• 📖 Documentation — improve docs, write tutorials
• 🐛 Bug reports — file issues with reproduction steps
• ⭐ Star the repo — if you find Quill useful, a star goes a long way!

---

📜 License

Apache 2.0

---

⭐ Star History

If Quill is useful to you, please consider giving it a star! Stars help others discover the project and motivate continued development.

---

Built with ❤️ by the Quill team · Inspired by OpenWork, DeerFlow, OpenClaw, Hermes Agent, Kimi Code, Codex, CrewAI, AutoGen, and awesome-harness-engineering