chatgpt-on-wechat

v2.0.9 Security

This release patches 1 CVE for security teams tracking exposure across their dependency inventory.

Published 12d AI Agents & Assistants

View tool

1 patched CVE

Read the diff → Tool health → What is this tool? →

This release patches 1 known CVE CVE-2023-4863 EPSS 93%

1 CVEs patched

Topics

ai ai-agent ai-agents chatgpt-on-wechat claude claude-code

+10 more

codex cowagent deepseek harness llm mcp multi-agent openai openclaw skills

ReleasePort's take

Moderate signal

editorial:auto 11d

Release 2.0.9 adds a model‑management page in the web console and introduces MCP protocol support, while fixing several UI and timer bugs.

Why it matters: Adds multi‑vendor model management (web) and open MCP integration for extensible tooling; resolves duplicate timer execution and clean‑exit handling – critical for developers and SREs managing diverse AI workloads.

Summary

AI summary

Updates 🛠 体验优化与修复, 🤖 模型新增与优化, and 🔒 部署与安全 across a mixed release.

Changes in this release

Type	Severity	Summary	CVE
Feature
Feature	Medium	Adds model management page in web console for multi‑vendor configuration, image, audio, vector models and search capabilities. Adds model management page in web console for multi‑vendor configuration, image, audio, vector models and search capabilities. Source: llm_adapter@2026-05-22 Confidence: high	—
Feature	Medium	Adds MCP (Model Context Protocol) support for open, pluggable tool ecosystem with native JSON‑RPC and stdio/sse transports. Adds MCP (Model Context Protocol) support for open, pluggable tool ecosystem with native JSON‑RPC and stdio/sse transports. Source: llm_adapter@2026-05-22 Confidence: high	—
Feature	Medium	Adds persistent browser login state reuse via ~/.cow/browser_profile and optional CDP mode for real Chrome integration. Adds persistent browser login state reuse via ~/.cow/browser_profile and optional CDP mode for real Chrome integration. Source: llm_adapter@2026-05-22 Confidence: high	—
Feature	Medium	Adds new models gpt-5.5, gemini-3.5-flash, qwen3.7-max, ernie-5.1 and optimizes DeepSeek V4 reasoning_effort parameter. Adds new models gpt-5.5, gemini-3.5-flash, qwen3.7-max, ernie-5.1 and optimizes DeepSeek V4 reasoning_effort parameter. Source: llm_adapter@2026-05-22 Confidence: high	—
Feature	Medium	Adds default binding of Web console to 127.0.0.1 and fully localizes frontend resources for offline use. Adds default binding of Web console to 127.0.0.1 and fully localizes frontend resources for offline use. Source: llm_adapter@2026-05-22 Confidence: high	—
Feature	Medium	Adds TTS support across multiple channels including WeChat, DingTalk, Feishu, and Enterprise WeCom. Adds TTS support across multiple channels including WeChat, DingTalk, Feishu, and Enterprise WeCom. Source: llm_adapter@2026-05-22 Confidence: low	—
Feature	Medium	Enhances log panel with level‑based highlighting and filtering capabilities. Enhances log panel with level‑based highlighting and filtering capabilities. Source: llm_adapter@2026-05-22 Confidence: low	—
Feature	Medium	Automatically opens Web console upon program startup. Automatically opens Web console upon program startup. Source: llm_adapter@2026-05-22 Confidence: low	—
Feature	Medium	Adds directory upload support in Web UI with Windows path validation. Adds directory upload support in Web UI with Windows path validation. Source: llm_adapter@2026-05-22 Confidence: low	—
Feature	Medium	Documents one‑click upgrade via `cow update` and manual code pull restart procedures. Documents one‑click upgrade via `cow update` and manual code pull restart procedures. Source: llm_adapter@2026-05-22 Confidence: low	—
Bugfix
Bugfix	Medium	Fixes timer task duplicate execution issue under certain conditions. Fixes timer task duplicate execution issue under certain conditions. Source: llm_adapter@2026-05-22 Confidence: high	—
Bugfix	Medium	Fixes single‑run timer task failure when timezone is specified. Fixes single‑run timer task failure when timezone is specified. Source: llm_adapter@2026-05-22 Confidence: high	—
Bugfix	Medium	Fixes noisy KeyboardInterrupt stack trace on Ctrl+C, enabling clean exit. Fixes noisy KeyboardInterrupt stack trace on Ctrl+C, enabling clean exit. Source: llm_adapter@2026-05-22 Confidence: low	—
Bugfix	Medium	Fixes failure to display failed tool calls after page refresh. Fixes failure to display failed tool calls after page refresh. Source: llm_adapter@2026-05-22 Confidence: low	—
Bugfix	Medium	Fixes Enterprise WeCom message delivery failure caused by illegal control characters. Fixes Enterprise WeCom message delivery failure caused by illegal control characters. Source: llm_adapter@2026-05-22 Confidence: low	—

Full changelog

🖥️ 新增模型管理

Web 控制台新增「模型」页面，按 模型厂商 + 模型能力 进行管理，支持对话、图像、语音、向量模型和搜索能力的配置：

多厂商配置：所有厂商的 API Key / API Base 在顶部统一维护，下方所有能力立即生效，无需重复填写
图像模型：图像理解与图像生成均可独立选择厂商和模型，未指定时跟随主模型自动选择
语音模型：语音识别和合成可独立配置，新增千问、智谱 ASR/TTS 模型
向量模型：支持配置 Embedding 模型（用于记忆及知识库检索），新增支持 OpenAI、通义、豆包、智谱等；切换模型后需执行 /memory rebuild-index 在线重建索引
搜索能力：联网搜索能力升级，支持博查、百度、智谱等多个厂商，自动模式下 Agent 可综合多来源搜索结果进行深度研究

🧩 MCP 协议支持

支持 MCP（Model Context Protocol） 协议，从固定工具集扩展为开放可插拔的工具生态，任何兼容 MCP 协议的服务均可作为工具直接接入 Agent。

原生 JSON-RPC 实现，零额外依赖，同时支持 stdio 和 sse 两种传输
兼容 Claude Desktop / Cursor 等主流风格的 mcpServers 配置，优先读取 ~/cow/mcp.json

相关文档：MCP 工具。Thanks @yangluxin613 (#2801)

🌐 浏览器登录态持久化

针对需要登录、有反爬机制的网站，浏览器工具支持登录一次后长期复用登录态，并允许接入用户自己的真实 Chrome 以通过指纹检测：

持久化用户配置（默认）：默认使用 ~/.cow/browser_profile 作为浏览器用户目录，登录一次后下次自动复用登录态
CDP 模式：通过 tools.browser.cdp_endpoint 接管真实 Chrome 浏览器，享有完整浏览器权限

相关文档：浏览器工具。Thanks @leafmove (#2809)

🤖 模型新增与优化

模型新增：gpt-5.5、gemini-3.5-flash、qwen3.7-max、ernie-5.1
模型优化：DeepSeek V4 支持 reasoning_effort 思考深度参数；修复 MiMo 等思考模型通过 OpenAI 兼容协议接入的问题

🔒 部署与安全

默认本机访问：Web 控制台 web_host 配置默认绑定 127.0.0.1，服务器部署时可手动设置为 0.0.0.0 并设置密码。Thanks @August829、@yidaozhongqing、@YLChen-007、@icysun
前端资源完全本地化：第三方 CSS / JS 全部本地分发，离线 / 内网环境也能正常加载控制台。Thanks @gitlayzer (#2816)

🛠 体验优化与修复

TTS 适配更多通道：Web对话、个人微信、飞书、钉钉、企微智能机器人均已支持回复语音，详情查看通道概览
日志面板增强：根据日志等级差异化高亮展示、支持根据等级筛选。Thanks @yangluxin613 (#2807)
Web 控制台自动启动：程序启动后自动打开 Web 控制台。Thanks @yangluxin613 (#2804)
Ctrl+C 干净退出：不再打印一长串 KeyboardInterrupt 堆栈。Thanks @yangluxin613 (#2806)
文件夹上传：Web 端支持目录上传，路径校验适配 Windows。Thanks @TryToMakeUsBetter (#2814)
修复定时任务在某些情况下重复执行的问题。Thanks @CNXudiandian (#2820)
修复定时任务带时区时单次任务不触发的问题。Thanks @AethericSpace
修复执行失败的工具调用在页面刷新后不显示的问题。Thanks @a1094174619 (#2822)
修复企微机器人消息中包含非法控制字符导致投递失败的问题。Thanks @Jacques-Zhao (#2810)

📦 升级方式

源码部署可执行 cow update 一键升级，或手动拉取代码后重启。详见更新升级文档。

发布日期：2026.05.22 | Full Changelog

Security Fixes

Web console `web_host` now defaults to binding only localhost (127.0.0.1) for safer server deployments

View diff on GitHub

Weekly OSS security release digest.

The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.

No spam, unsubscribe anytime.

Share this release

Share on X Share on Bluesky

Track chatgpt-on-wechat

Get notified when new releases ship.

About chatgpt-on-wechat

CowAgent是基于大模型的超级AI助理，能主动思考和任务规划、访问操作系统和外部资源、创造和执行Skills、通过长期记忆和知识库不断成长，比OpenClaw更轻量和便捷。同时支持微信、飞书、钉钉、企微、QQ、公众号、网页等接入，可选择OpenAI/Claude/Gemini/DeepSeek/ Qwen/GLM/Kimi/LinkAI，能处理文本、语音、图片和文件，可快速搭建个人AI助理和企业数字员工。

All releases →