Baize MCP 0.1.4

Baize MCP v0.1.4

中文

Baize MCP v0.1.4 在命令计划工作流之外补齐了观察、诊断与远程任务闭环，让 AI 客户端可以在白泽账号权限范围内完成"了解状态 → 定位问题 → 执行/跟踪任务"的完整流程，同时保持结果字段、数量和文本长度的有界输出。

主要变化

• 增加运行概览工具，读取当前账号可见范围内的平台运行摘要和有限数量的重点异常节点；缓存缺失、异常列表为空、分区失败均显式标记。
• 增加节点有界观察工具，按健康、指标、进程、存储、Docker、Nginx、主机画像状态和控制面等维度读取单一节点；敏感正文、凭据、环境变量和完整历史不返回。
• 增加固定语义观察工具：告警列表、资产清单与到期、证书监控状态、定时任务、日志、Nginx、Runbook、安全暴露面、订阅状态和组件版本状态；各视图只返回完成判断所需字段，关联标识、地址和原始路径均已脱敏。
• 增加只读运行态诊断：可以为单个节点发起诊断探测、查询状态与证据数量，并在脱敏后读取诊断 AI 上下文。
• 补齐远程任务闭环：可以通过直接任务入口创建模板任务或服务端允许的精确自定义命令，把等待中的任务推入执行链路，在用户明确要求后按目标与游标读取有界输出，并继续支持查询进度和请求取消。
• 增加告警确认/解决工具；服务端负责权限、状态规则和审计，成功后需再次查询确认最终状态。
• 增加工作流状态工具：读取本机单人/多人工作流模式和服务端审批策略摘要；无策略查看权限时仍返回本地模式并显式标记。
• 增加命令计划取消工具；错误信息保留稳定的原因、可重试标记、消息键和下一步动作键，不再透传底层错误原文。
• AI 接入说明覆盖 DeepSeek Harness（DSH）：接入安装器会自动注册，也给出用户插件层的标准配置。
• 最低兼容白泽版本提升为 0.2.2，以使用直接任务派发等新发布的稳定行为。

兼容性与升级

• 最低兼容白泽版本：0.2.2。
• 传输方式：stdio。
• 支持 Linux、macOS、Windows 的 amd64 与 arm64 发布包，每个包内含可执行文件、启动校验文件、项目许可证、中英文说明和第三方许可证清单。
• 已安装用户可从此前获取的 baize 公开入口运行 AI 接入升级器；升级会校验发布归档和可执行文件，并保留本机 profile 与系统凭据。升级前退出正在使用 MCP 的 AI 客户端，完成后重新打开。
• 手动下载时请在 GitHub Releases 核对 SHA256SUMS；每个发布包内含 baize-mcp.sha256 完整性文件，程序启动时会自动校验可执行文件，发现文件损坏或安装不完整即拒绝运行。
• 白泽权限、审批策略、风险确认和审计继续由服务端处理；MCP 不复制这些控制，审批通过也不代表计划会自动执行。

完整安装和使用说明见 Baize 公开入口 的 AI 接入与远程任务指南。

English

Baize MCP v0.1.4 completes the loop beyond the command-plan workflow: observation, diagnosis, and remote-task execution, so an AI client can cover "understand state → locate issues → execute and track tasks" within the signed-in Baize account's permissions while keeping bounded fields, counts, and text lengths.

Highlights

• Adds a runtime overview tool with the account-scoped platform summary and a bounded set of highlighted abnormal nodes; missing caches, an empty abnormal list, and failed sections are marked explicitly.
• Adds bounded agent observation for one node across health, metrics, processes, storage, Docker, Nginx, host-profile, and control-plane views; sensitive bodies, credentials, environment values, and complete histories are excluded.
• Adds fixed semantic observation tools: alert lists, asset inventory and expiry, certificate status, scheduled tasks, logs, Nginx, Runbooks, security exposure, subscription status, and component version status; each view returns only the fields needed for the current decision, with correlation identifiers, addresses, and raw paths redacted.
• Adds read-only runtime diagnosis: start a probe for one node, query its status and evidence counts, and read the privacy-reduced diagnosis AI context.
• Closes the remote-task loop: create template tasks or exact custom commands allowed by Baize through the direct-task entry, push pending tasks into the execution chain, read bounded output by target and cursor after the user explicitly asks for it, and keep progress lookup and cancellation.
• Adds alert acknowledgement/resolution; Baize handles permissions, state rules, and audit, and callers must query the alert again to confirm the final status.
• Adds a workflow-status tool reading the local single-user/multi-user mode and a minimal server approval-policy summary; when policies are not viewable, the local mode is still returned with an explicit marker.
• Adds command-plan cancellation, and keeps stable reason, retryable, message-key, and next-action fields in errors instead of exposing raw backend messages.
• AI access documentation now covers DeepSeek Harness (DSH): the installer registers it automatically, and the standard patch-layer configuration is documented.
• The minimum compatible Baize version is raised to 0.2.2 for the newly released stable direct-task dispatch and related behavior.

Compatibility and upgrade

• Minimum compatible Baize version: 0.2.2.
• Transport: stdio.
• Release archives are available for Linux, macOS, and Windows on amd64 and arm64; each contains the executable, startup integrity file, project license, bilingual guides, and third-party license notices.
• Existing users can run the AI access upgrader from the previously obtained baize public entry. The upgrader verifies the release archive and executable, and preserves the local profile and operating-system credential store. Close AI clients using MCP before upgrading, then reopen them afterward.
• For manual downloads, verify the archive against SHA256SUMS on GitHub Releases; every archive also ships a baize-mcp.sha256 integrity file and the executable verifies itself at startup, refusing to run when the file is corrupted or the installation is incomplete.
• Baize permissions, approval policy, risk confirmation, and audit remain server-side; MCP does not replicate those controls, and approval does not cause a plan to execute automatically.

See the Baize public entry and its AI Access and Remote Task Guide for complete installation and usage instructions.