v2.3.2 · Local Autonomous Agent

COGITO

AGENT

Think Continuously · Act Autonomously · Stay Private

持续思考 · 自主行动 · 数据私有

"Cogito, ergo sum — I think, therefore I am. CogitoAgent is not just a tool; it is your autonomous thinking partner running entirely in your local environment."

— CogitoAgent v2.3.2

GitHubGitHub GiteeGitee Issues反馈

Your Local AI Autonomous Agent您的本地 AI 自主智能体

CogitoAgent is a locally-run autonomous AI agent that integrates file management, knowledge mining, system operations, code execution, and web connectivity. It runs directly within the working directory you configure — with no need to upload files to third-party servers. CogitoAgent 是一款本地运行的自主 AI 智能体，集成了文件管理、知识挖掘、系统操作、代码执行、网络搜索等功能。它直接在您配置的工作目录内运行，无需将文件上传至第三方服务器。

Unlike traditional chatbots, CogitoAgent possesses the ability to Think Continuously, Explore Autonomously, and Execute Tools, capable of proactively discovering and organizing your local assets in the background. 与传统聊天机器人不同，CogitoAgent 拥有 持续思考、自主探索 和 工具执行 的能力，可以在后台主动发现和整理您的本地资源。

Continuous Thinking持续思考

Event-driven, thinks continuously as events arrive事件驱动，随事件持续思考

Privacy First隐私优先

All data stays local, no uploads数据完全本地，无需上传

Rich Toolset丰富工具集

28 modules: files, code, Git, OCR, DB, GIS, Bio, Med, Chem, Finance...28个模块：文件、代码、Git、OCR、数据库、GIS、生物、医学、化学、金融等

Persistent Memory持久记忆

Long-term storage, cross-session learning长期记忆存储，跨会话学习

cogito-agent

$ npm start

[think] Scanning workspace...

[think] Found 3 pending tasks

[think] Fetching latest updates...

✓ Planning complete, awaiting instructions

[think] Loading persona: Default — your intelligent assistant

✓ Agent ready. Type '/help' for available commands.

🤖 Hello! I'm CogitoAgent, your local autonomous AI assistant 🧠

✨ I'm continuously thinking and exploring your workspace to help organize, analyze, and create.

💡 What can I do for you today? Try '/help' to see all available commands!

Desktop Preview桌面预览

200+

Tool Functions工具函数

Tool Modules工具模块

Preset Personas预设人设

Quick Commands快捷指令

Supported Models支持模型

✓OpenAI (GPT-4o)

✓Moark

✓Anthropic (Claude)

✓Google (Gemini)

✓Ollama (Local)

✓OpenAI-Compatible API

Tool Modules工具模块

All tool functions are managed uniformly by registry.ts and called using the [TOOL] functionName(args) [/TOOL] format. Tools are loaded on-demand by category. 所有工具函数由 registry.ts 统一管理，使用 [TOOL] functionName(args) [/TOOL] 格式调用。工具按分类按需加载。

File Operations文件操作

ls, read, create, copy, mkdir, delete, rename

Web Tools网页工具

search, browse, fetchPage, searchOnEngine

Code Execution代码执行

JavaScript, Python, formatCode, executeFile

Git VCSGit 版本控制

commit, push, branch, checkout, diff, stash

Task Management任务管理

create, track, split, complete, stats

Memory System记忆系统

store, search, recall, stats, related

Database数据库

SQLite queries, transactions, schema

Image Recognition (OCR)图像识别 (OCR)

text extraction, batch processing

Vision Analysis视觉分析

image description, content understanding

Email邮件功能

text, HTML, attachments, templates

Scheduled Tasks定时任务

cron-style scheduling, toggle

System Monitor系统监控

CPU, memory, disk, processes

WeChat Channel微信通道

QR login, receive/send messages, dedicated session

Browser Automation浏览器自动化

click, fill, screenshot, navigate

GIS & Maps地理信息系统

geocode, reverse geocode, spatial query, routing

Live Interactive Demos实时交互演示

Experience CogitoAgent's three modes — Desktop, Dashboard, and Monitor — all running live in your browser with simulated data. 体验 CogitoAgent 的三种模式 — 桌面端、Dashboard 和监控面板，全部在浏览器中实时模拟运行。

Desktop Mode桌面模式 LIVE实时

Dashboard ModeDashboard 模式 LIVE实时

Monitor Panel监控面板 LIVE实时

Vision & OCR视觉与 OCR

CogitoAgent integrates advanced computer vision capabilities, enabling image analysis, OCR text extraction, and visual understanding. CogitoAgent 集成了先进的计算机视觉能力，支持图像分析、OCR 文本提取和视觉理解。

Vision Analysis视觉分析

Analyze images for content description, object detection, scene understanding, and more using multi-modal AI models. 使用多模态 AI 模型分析图像内容描述、物体检测、场景理解等。

Web Search Preview网页搜索预览

Search the web, browse pages, and extract structured information from any public URL. 搜索网页、浏览页面，从任意公开 URL 提取结构化信息。

        // Example: Vision analysis

        [TOOL] vision("screenshot.png", "Describe this image") [/TOOL]

        // Example: OCR text extraction

        [TOOL] ocr("document.png") [/TOOL]

        // Example: Batch OCR

        [TOOL] ocrBatch("img1.png, img2.png") [/TOOL]

Architecture架构设计

The Agent core runs an event-driven thinking loop: on each user message or tool result it builds context → calls the API → parses output → executes tools → and immediately continues. No fixed polling interval — turns advance as fast as the model and tools allow. Agent 核心运行事件驱动的思考循环：每收到一条用户消息或工具结果，便构建上下文 → 调用 API → 解析输出 → 执行工具 → 并立即推进下一轮。无固定轮询间隔，循环速度取决于模型与工具的实际耗时。

🧠

Thinking Loop思考循环

3s interval auto-trigger3秒间隔自动触发

⚙️

State Machine状态机

THINKING → AWAIT_INPUT → AWAIT_CONFIRM思考 → 等待输入 → 等待确认

🔌

Plugin System插件系统

Dynamic tool loading动态工具加载

🛡️

Sandbox Security沙箱安全

isolated-vm process isolationisolated-vm 进程隔离

📁

Multi-Session多会话管理

Persistent storage, context compression持久化存储，上下文压缩

🌐

MCP ProtocolMCP 协议

Model Context Protocol support模型上下文协议支持

📋

Tool Registry工具注册表

Centralized tool management集中式工具管理

🔔

Event Bus事件总线

Pub/sub inter-module communication发布/订阅模块间通信

🧑

Persona System人设系统

25 preset role profiles25 个预设角色档案

🌉

Streaming Bridge流式桥接

WebSocket real-time relayWebSocket 实时中继

Technical Details技术细节

🔄

State Machine状态机

Three core states managed by state.ts with strict transitions. 由 state.ts 管理的三个核心状态，具有严格的状态转换。

            // State transitions

            AWAITING_INPUT → THINKING (user input)

            THINKING → AWAITING_CONFIRMATION (dangerous tool)

            THINKING → AWAITING_INPUT (response ready)

            AWAITING_CONFIRMATION → THINKING (confirmed)

🔌

Tool Registry工具注册表

Centralized tool management with metadata including category, parameters, danger level, and execution handler. 集中式工具管理，包含分类、参数、危险等级和执行句柄等元数据。

            // Tool registration format

            name: 'functionName',

            description: 'description',

            parameters: { /* JSON Schema */ },

            category: 'file/code/git/...',

            dangerLevel: 'none/low/medium/high'

🛡️

Security Sandbox安全沙箱

Process-level isolation using isolated-vm with deep-freezing of built-in objects and resource limits. 使用 isolated-vm 实现进程级隔离，内置对象深度冻结，资源使用受限。

            // Sandbox security features

            ✅ Deep freeze of built-in objects

            ✅ No file system access

            ✅ No network access

            ✅ Execution timeout limit

            ✅ Memory usage limit

📦

Context Compression上下文压缩

Auto-compression when conversation exceeds max length. Uses LLM to summarize early messages while keeping recent context. 当对话超过最大长度时自动压缩。使用 LLM 总结早期消息，保留最近上下文。

            // Compression strategy

            maxHistoryLength: 100 (messages)

            compressThreshold: 80% of max

            keepRecent: 20 latest messages

            summarize: Early messages → summary

📈

Tracing & Retry追踪与重试

Lightweight observability with tracing.ts and reliable network requests with retry.ts using exponential backoff. 轻量级可观测性（tracing.ts）和可靠网络请求（retry.ts），支持指数退避策略。

            // Retry configuration

            maxAttempts: 3

            delay: 1000ms

            backoff: 'exponential'

            retryOn: [Error, Timeout]

🌐

MCP ProtocolMCP 协议

Model Context Protocol support for integration with other AI clients. Exposes tools via JSON-RPC. 支持模型上下文协议（MCP），可与其他 AI 客户端集成。通过 JSON-RPC 暴露工具。

            // MCP methods

            tools/list - List all tools

            tools/call - Invoke tool

            resources/list - List resources

            resources/read - Read resource

Thinking Cycle Workflow思考循环流程

Perceive Input感知输入

→

Build Prompt构建提示

→

Call LLM调用 LLM

→

Parse Response解析响应

→

Execute Tools执行工具

→

Update State更新状态

↻

Continues until the response is ready; turns advance immediately as tool results arrive 持续运行直到响应就绪；工具结果返回后立即推进下一轮

Project Structure项目结构

          // Core files and their responsibilities

          src/

          ├── agent/ // Agent core modules

          │   ├── Agent.ts // Thinking loop controller

          │   ├── state.ts // State machine management

          │   ├── registry.ts // Tool registration & discovery

          │   ├── session.ts // Multi-session management

          │   ├── commands.ts // Command handling (/help, /status, etc.)

          │   ├── stats.ts // Usage statistics & metrics

          │   ├── tracing.ts // Lightweight observability

          │   ├── cluster-commands.ts // Sub-agent commands

          │   ├── mcp.ts // MCP Server protocol

          │   ├── plugin.ts // Plugin system

          │   └── tools/ // 28+ tool modules

          ├── api/ // API layer (client, models, webSearch)

          ├── io/ // Terminal, Logger, WebSocket

          ├── config.ts // Configuration management

          └── index.ts // Application entry

          electron/ // Desktop mode (Electron)

          personas/ // 25 preset personas

          data/ // Runtime data (auto-created)

Getting Started快速开始

CogitoAgent offers multiple operation modes to suit different environments and preferences. CogitoAgent 提供多种运行模式以适应不同环境和偏好。

🖥️ Desktop Mode🖥️ 桌面模式

Electron desktop window + terminal Agent, communicating via WebSocket (port 9527). Features a semi-transparent frosted glass dialog for seamless interaction. Electron 桌面窗口 + 终端 Agent，通过 WebSocket（端口 9527）通信。半透明毛玻璃对话框，实现无缝交互。

npm start

📊 Dashboard Mode📊 仪表盘模式

Full-window dashboard with session management, tool usage statistics, and real-time monitoring. This is the default mode of `npm start`. 全屏仪表盘，包含会话管理、工具使用统计和实时监控，即 npm start 的默认模式。

npm start

💻 CLI Mode💻 命令行模式

Pure command-line mode without Electron or WebSocket. Ideal for servers, remote connections, or resource-constrained environments. 纯命令行模式，无 Electron 或 WebSocket。适用于服务器、远程连接或资源受限环境。

npm run cli

🚀 Setup Wizard🚀 配置向导

Interactive setup wizard for first-time configuration: API settings, workspace path, and persona selection. 交互式配置向导，首次运行时引导完成：API 设置、工作区路径和人设选择。

npm start

On first run (npm start), the setup wizard will guide you through API configuration, workspace setup, and persona selection. 首次运行时（npm start），配置向导将引导您完成 API 配置、工作区设置和人设选择。

Quick Commands快捷指令

ENTER Interrupt thinking, enter command打断思考，输入指令

exit Exit program退出程序

/help Show help显示帮助

/status View status查看状态

/persona Switch persona切换人设

/personas List all personas列出所有人设

/tools List all tools查看所有工具

/config Show configuration查看配置

/sessions List all sessions列出所有会话

/new Create new session创建新会话

/switch Switch session切换会话

/clear Clear conversation清除对话

/wechat/status WeChat channel status微信通道状态

/wechat/login Login via QR code扫码登录微信

/wechat/logout Logout from WeChat退出微信登录

/spawn Create sub-agent创建子智能体

/agents List sub-agents列出子智能体

/delegate Delegate task委派任务

Ready to Get Started?准备好开始了吗？

Choose your preferred mode and let CogitoAgent work for you. 选择您喜欢的模式，让 CogitoAgent 为您工作。

View on GitHub在 GitHub 查看 → Mirror: Gitee镜像：Gitee Learn More了解更多