Castor AI 工作流
经 AI Skill Hub 精选评估，Castor AI 工作流 获评「推荐使用」。这款Agent工作流在功能完整性、社区活跃度和易用性方面表现出色，AI 评分 7.5 分，适合有一定技术背景的用户使用。
📚 深度解析
Castor AI 工作流 工作流的设计遵循"最小配置，最大复用"原则：核心逻辑已经封装好，用户只需配置自己的 API Key 和业务参数即可快速上手。工作流内置错误处理和重试机制，在网络波动或 API 限速等情况下仍能稳定运行，适合作为生产环境的自动化基础设施。
在实际部署时，建议先在测试环境中运行 3-5 次，验证各个环节的输出结果符合预期，再部署到生产环境。AI Skill Hub 评分 7.5 分，是同类 Agent 工作流中的精选推荐。
📋 工具概览
Castor AI 工作流 是一套完整的 AI Agent 自动化工作流方案。通过可视化的节点编排，将复杂的多步骤任务拆解为清晰的自动化流程，实现全程无人值守的智能处理。支持与数百种外部服务和 API 无缝集成，适合构建数据处理管线、业务自动化和 AI 辅助决策系统。
📖 中文文档
Castor AI 工作流 是一套完整的 AI Agent 自动化工作流方案。通过可视化的节点编排，将复杂的多步骤任务拆解为清晰的自动化流程，实现全程无人值守的智能处理。支持与数百种外部服务和 API 无缝集成，适合构建数据处理管线、业务自动化和 AI 辅助决策系统。
- 可视化 Agent 工作流编排，无需编写复杂代码
- 支持多步骤自动化任务链，实现全流程无人值守
- 与外部 API、数据库和第三方服务无缝集成
- 内置错误处理与自动重试机制，保障稳定运行
- 提供可复用的自动化模板，快速在同类场景部署
- 自动化日常重复性工作，将精力集中于创造性任务
- 构建数据采集 → 处理 → 输出的完整自动化管线
- 实现跨平台、跨系统的数据流转和业务协同
# 方式一：pip 安装（推荐）
pip install castor
# 方式二：虚拟环境安装（推荐生产环境）
python -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install castor
# 方式三：从源码安装（获取最新功能）
git clone https://github.com/deepfounder-ai/castor
cd castor
pip install -e .
# 验证安装
python -c "import castor; print('安装成功')"
- 访问 GitHub 仓库获取工作流文件
- 在对应平台（Dify / Flowise / Make 等）中找到「导入工作流」功能
- 上传工作流文件
- 按照提示配置必要的环境变量和 API Key
- 运行测试确认流程正常后投入使用
# 命令行使用
castor --help
# 基本用法
castor input_file -o output_file
# Python 代码中调用
import castor
# 示例
result = castor.process("input")
print(result)
# castor 配置文件示例（config.yml） app: name: "castor" debug: false log_level: "INFO" # 运行时指定配置文件 castor --config config.yml # 或通过环境变量配置 export CASTOR_API_KEY="your-key" export CASTOR_OUTPUT_DIR="./output"
简介
<p align="center"> <pre> ██████╗ █████╗ ███████╗████████╗ ██████╗ ██████╗ ██╔════╝██╔══██╗██╔════╝╚══██╔══╝██╔═══██╗██╔══██╗ ██║ ███████║███████╗ ██║ ██║ ██║██████╔╝ ██║ ██╔══██║╚════██║ ██║ ██║ ██║██╔══██╗ ╚██████╗██║ ██║███████║ ██║ ╚██████╔╝██║ ██║ ╚═════╝╚═╝ ╚═╝╚══════╝ ╚═╝ ╚═════╝ ╚═╝ ╚═╝ </pre> </p>
Business-oriented AI agent
<p align="center"> Self-hosted AI agent ready to drop into business workflows. Bring any OpenAI-compatible LLM — Azure OpenAI, AWS Bedrock, OpenAI, Groq, OpenRouter, or a local model on your own hardware. Your data, your provider, your rules. </p>
<p align="center"> <a href="#quick-start">Quick Start</a> • <a href="#interfaces">Interfaces</a> • <a href="docs/README.md"><b>Documentation</b></a> • <a href="#tools">Tools</a> • <a href="#skills">Skills</a> • <a href="#telegram-bot">Telegram</a> • <a href="#diagnostics">Doctor</a> </p>
<p align="center"> <img src="https://img.shields.io/badge/version-0.23.2-blue" alt="version"> <img src="https://img.shields.io/badge/python-3.11+-green" alt="python"> <img src="https://img.shields.io/badge/platform-linux%20%7C%20macos%20%7C%20windows-lightgrey" alt="platform"> <img src="https://img.shields.io/badge/license-MIT-orange" alt="license"> <a href="https://t.me/castor_ai"><img src="https://img.shields.io/badge/community-Telegram-blue?logo=telegram" alt="Telegram"></a> </p>
---
Features
Castor's design principle: the system around the LLM does the heavy lifting, so the agent stays reliable on small local models and cheap on large hosted ones. Each feature below has a deep-dive guide in docs/.
Tool Search — a meta-tool architecture that keeps the prompt lean. Only ~8 core tools load by default (memory_search, memory_save, read_file, write_file, shell, http_request, spawn_task, tool_search); the model calls tool_search("browser") / "schedule" / "secret" / … to activate the rest on demand. Saves ~75% of the tokens a flat 49-tool list would burn.
Memory & Knowledge Graph — 3-layer system in one Qdrant collection: raw facts (saved instantly) → entities with typed relations → wiki summaries (both built by a nightly synthesis job). Hybrid retrieval fuses dense (FastEmbed MiniLM, 384d, 50+ languages) + sparse (SPLADE++) + BM25 via RRF. Thread-isolated, auto-chunked, secret-scrubbed. Interactive force-directed graph in the Web UI. → docs/MEMORY.md
Knowledge ingest — 50+ formats via Microsoft MarkItDown: PDF / DOCX / PPTX / XLSX / EPUB / HTML / code / data / images. Drop files, paste a URL, or scan a folder. Chunked, embedded, and queued for entity + wiki synthesis. → docs/KNOWLEDGE.md
Skills — pluggable single-file Python modules. Nine built in (browser, canvas, serial_port, mcp_manager, skill_creator, soul_editor, notes, timer, weather); create new ones from chat (skill_creator runs a plan→code→validate pipeline) or import from the agentskills.io spec. → docs/SKILLS.md · docs/SKILLS_IMPORT.md
Browser — Playwright + Chromium. Navigate, read, click, fill forms, screenshot. Headless by default; visible mode for logged-in sessions and OAuth flows. → docs/BROWSER.md
Hardware — the serial_port skill talks USB-serial / RS-232 / RS-485 to scales, barcode/RFID readers, GPS, label & receipt printers, PLCs (Modbus RTU), VFDs, and sensors. Cross-platform via pyserial. Actuator writes are gated behind an explicit confirm=true with a hex preview. → docs/HARDWARE.md
Canvas — render model-supplied HTML in a sandboxed 480px side panel: blocking forms that return submitted data, saveable dashboards, throwaway mockups. Iframe is sandbox="allow-scripts allow-forms" with no allow-same-origin, so generated HTML can't read parent state. → docs/CANVAS.md
Routines — scheduled tasks that live as chat threads: each firing appends a turn, and corrections you add between runs become context for the next. Natural schedule syntax (every 2h, weekdays 09:00, mon,wed,fri 14:30). Per-routine USD budget caps. → docs/ROUTINES.md
Goals — long-running autonomous tasks. A durable SQLite queue + worker daemon survives disconnects and restarts; an orchestrator breaks the goal into subtasks, dispatches specialized subagents, and an acceptance gate validates deliverables before marking done. → docs/GOALS.md
MCP — connect external Model Context Protocol tool servers (stdio or HTTP). Tools surface as mcp__server__tool and flow through tool_search. Manage via chat (mcp_manager) or Settings. → docs/MCP.md
Providers — any OpenAI-compatible endpoint (LM Studio, Ollama, OpenAI, OpenRouter, Groq, Together, DeepSeek, + more) plus a native Anthropic adapter for prompt caching & thinking budgets. Switch per-thread via /model or Settings. → docs/PROVIDERS.md
Voice & Camera — live voice mode (VAD → STT → LLM → TTS → auto-listen), local or API STT, multiple TTS backends; camera capture via browser PiP or OpenCV. → docs/VOICE.md · docs/CAMERA.md
Personality (Soul) — 8 adjustable traits (humor, honesty, curiosity, brevity, formality, proactivity, empathy, creativity) plus custom traits, agent name, and language. Edit via /soul, Settings, or chat. → docs/SOUL.md
Cost tracking — every LLM call records tokens + USD by thread, source, model, and provider, with LiteLLM-backed pricing. Surfaced in the Web UI. → docs/COST_TRACKING.md
The reliability internals that keep all of this working on small models — JSON repair, anti-hedge nudging, self-check, loop detection, compaction, auto-resume — are described under Engineering around the LLM above.
Prerequisites
- Python 3.11+
- An LLM endpoint — pick one:
- Hosted (any OpenAI-compatible API): Azure OpenAI, AWS Bedrock, OpenAI, Groq, OpenRouter, DeepSeek, Together. Set
CASTOR_LLM_URL+CASTOR_LLM_KEYand you're done. - Local (data stays on-prem): LM Studio or Ollama with any tool-capable model. Qwen 9B / Gemma 4B work well on a single consumer GPU; bigger models if you have the hardware.
- Embeddings: FastEmbed (ONNX, local, CPU) — multilingual-MiniLM (384d, 50+ languages) + SPLADE++. Runs comfortably on a laptop without a GPU.
System requirements
For hosted-LLM deployments, Castor itself is light — any modern laptop or small VM works (the agent process is ~300MB resident, plus Qdrant on disk for memory).
For local-LLM deployments where the model runs on the same machine:
| Component | Minimum | Recommended |
|---|---|---|
| GPU | 4GB VRAM (4B Q4) | 8GB VRAM (9B Q4_K_M) or larger |
| RAM | 8GB | 16GB |
| Storage | 10GB | 20GB (models + memory) |
Works on: gaming laptops, desktop GPUs (RTX 3060+), Mac M1+ (via Ollama), Linux servers.
Install
Runs natively on Linux, macOS (Intel & Apple Silicon) and Windows 10/11 — single pip install -e . pulls every runtime dep (including MarkItDown, python-docx/pptx, openpyxl, pdfminer.six, pypdf, fastembed, qdrant-client, uvicorn).
🐧 Linux / 🍎 macOS — one-line
curl -fsSL https://raw.githubusercontent.com/deepfounder-ai/castor/main/install.sh | bashThis clones the repo, creates a venv, installs everything, verifies critical deps, pre-downloads the embedding model, and drops castor on your $PATH.
🪟 Windows
git clone https://github.com/deepfounder-ai/castor.git
cd castor
setup.batOn Windows shell commands are routed through Git Bash (auto-detected at install time — install Git for Windows if missing). Falls back to cmd.exe if not found.
Manual (any platform)
```bash git clone https://github.com/deepfounder-ai/castor.git cd castor
Install package + all runtime deps
pip install -e .
Docker
docker compose upLM Studio / Ollama should be running on the host. Persistent data in ./data/.
A module-by-module map of the codebase lives in ARCHITECTURE.md.
Quick Start
Create venv
python3 -m venv .venv # or python -m venv .venv on Windows source .venv/bin/activate # macOS/Linux
.venv\Scripts\activate # Windows PowerShell / cmd
Config
Environment variables:
CASTOR_LLM_URL=http://localhost:1234/v1 # LLM server URL
CASTOR_LLM_MODEL=qwen/qwen3.5-9b # Model name
CASTOR_LLM_KEY=lm-studio # API key
CASTOR_DB_PATH=~/.castor/castor.db # Database path
CASTOR_DATA_DIR=~/.castor # Where threads / memory / uploads live
CASTOR_QDRANT_MODE=disk # memory | disk | server
CASTOR_PASSWORD= # Web UI password (shows login modal if set)
CASTOR_STT_DEVICE=cpu # STT inference device (cpu | cuda)Everything else (30+ knobs — context_budget, rag_chunk_size, synthesis_time, tts_api_url, etc.) lives in Settings → Advanced → Settings and persists in SQLite.
Interfaces
- Web UI —
castor --web(add--ssl --port 7861for mic/camera). Single-file SPA, zero runtime JS deps. Chat, memory browser, scheduler, presets, settings, knowledge graph, canvas panel, live voice mode. - Terminal —
castor. Rich-formatted chat with 20+ slash commands (/soul,/skills,/memory,/model,/cron,/doctor, …). - Telegram — full mobile access: streaming replies, slash commands, topic-to-thread mapping, image vision. Setup → docs/TELEGRAM.md.
高质量的开源AI工作流项目，值得关注
- 构建多智能体协作系统的 Agent 开发者
- Agent 任务先做 dry-run 验证工具调用链，再开启自主执行
- API key 直接提交到 git 仓库（请用 .env 并加入 .gitignore）
- Python 依赖冲突：建议用 venv / uv 隔离环境
- 云端托管：可放在 Vercel / Railway / Fly.io 等 PaaS 平台
⚡ 核心功能
- 可视化 Agent 工作流编排，无需编写复杂代码
- 支持多步骤自动化任务链，实现全流程无人值守
- 与外部 API、数据库和第三方服务无缝集成
- 内置错误处理与自动重试机制，保障稳定运行
- 提供可复用的自动化模板，快速在同类场景部署
- 构建多智能体协作系统的 Agent 开发者
- Agent 任务先做 dry-run 验证工具调用链，再开启自主执行
- API key 直接提交到 git 仓库（请用 .env 并加入 .gitignore）
- Python 依赖冲突：建议用 venv / uv 隔离环境
👥 适合人群
🎯 使用场景
- 自动化日常重复性工作，将精力集中于创造性任务
- 构建数据采集 → 处理 → 输出的完整自动化管线
- 实现跨平台、跨系统的数据流转和业务协同
⚖️ 优点与不足
- +MIT 协议，可免费商用
- +大幅减少重复性人工操作
- +可视化流程，清晰直观
- +可扩展性强，支持复杂场景
- −初始配置和调试需投入一定时间
- −强依赖外部服务的稳定性
- −复杂场景需具备一定技术基础
AI Skill Hub 为第三方内容聚合平台，本页面信息基于公开数据整理，不对工具功能和质量作任何法律背书。
建议在沙箱或测试环境中充分验证后，再部署至生产环境，并做好必要的安全评估。
✅ MIT 协议 — 最宽松的开源协议之一，可自由商用、修改、分发，仅需保留版权声明。
🔗 相关工具推荐
❓ 常见问题 FAQ
AI Skill Hub 点评：Castor AI 工作流 的核心功能完整，质量良好。对于自动化工程师和运维人员来说，这是一个值得纳入个人工具库的选择。建议先在非生产环境试用，再逐步推广。
| 原始名称 | castor |
| 原始描述 | 开源AI工作流：A self-hosted AI agent built to drop into business workflows: customer ops, inte。⭐48 · Python |
| Topics | aiagentpythonworkflow |
| GitHub | https://github.com/deepfounder-ai/castor |
| License | MIT |
| 语言 | Python |
收录时间：2026-05-31 · 更新时间：2026-06-01 · License：MIT · AI Skill Hub 不对第三方内容的准确性作法律背书。