opus-advisor-mcp
opus-advisor-mcp
一个 MCP 服务器,允许 Claude Code 在任务中途咨询 Opus 作为战略顾问。在 Sonnet 或 Haiku 上运行会话,并按需将复杂决策升级给 Opus——使用您现有的 Claude Code 订阅。
灵感来自 Anthropic 的顾问策略。
工作原理
┌─────────────────────────────────────────────┐
│ Claude Code (Sonnet) │
│ │
│ "I need to decide on the DB schema..." │
│ │ │
│ ▼ │
│ calls consult_opus MCP tool │
│ │ │
└────────┼────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────┐
│ opus-advisor MCP server │
│ │
│ 1. Reads prior consultation history │
│ 2. Reads requested files from disk │
│ 3. Pipes prompt to: claude -p --model opus │
│ 4. Logs advice to advisor-log.md │
│ 5. Returns advice to Sonnet │
└─────────────────────────────────────────────┘无需 API 密钥。该服务器通过调用 claude CLI 来运行,该 CLI 使用您现有的身份验证。
Related MCP server: codex-bridge
安装
npm install -g opus-advisor-mcp或者克隆并在本地构建:
git clone https://github.com/Divinci-AI/opus-advisor-mcp.git
cd opus-advisor-mcp
npm install
npm run build配置
添加到您项目的 .mcp.json 或 ~/.claude/.mcp.json 中:
{
"mcpServers": {
"opus-advisor": {
"command": "opus-advisor",
"timeout": 180000
}
}
}如果是在本地安装(而非全局安装):
{
"mcpServers": {
"opus-advisor": {
"command": "node",
"args": ["/path/to/opus-advisor-mcp/dist/index.js"],
"timeout": 180000
}
}
}添加配置后重启 Claude Code。
工具
consult_opus
咨询 Opus 以获取战略建议。
参数 | 类型 | 默认值 | 描述 | ||
| string | 必填 | 您需要咨询的问题或难题 | ||
| string | 可选 | 额外的上下文、约束或背景 | ||
| string[] | 可选 | 作为代码上下文包含的文件路径(相对于项目根目录) | ||
|
|
|
|
| Opus 的推理努力程度 |
| boolean |
| 包含之前的咨询历史以保持连贯性 |
示例:
{
"question": "Is this database migration safe under concurrent writes?",
"files": ["src/db/migration-042.ts", "src/db/schema.ts"],
"effort": "high"
}read_advisor_log
读取之前调用的咨询日志。
参数 | 类型 | 描述 |
| number | 要返回的最近咨询次数(省略则返回全部) |
read_advisor_meta
读取结构化元数据(延迟、令牌计数、努力程度)。
参数 | 类型 | 描述 |
| number | 要返回的最近条目数(省略则返回全部) |
clear_advisor_log
清除咨询日志和元数据以重新开始。
功能
无需 API 密钥 — 通过
claudeCLI 使用您现有的 Claude Code 订阅按项目记录日志 — 咨询历史存储在每个项目的
~/.opus-advisor/<project>-<hash>/中代码感知上下文 — 直接传递文件路径;服务器会读取并将其作为标记的代码块注入
咨询连贯性 — 之前的建议会作为上下文反馈,以便 Opus 可以在之前的决策基础上进行构建
令牌感知历史 — 历史记录受条目数 (5) 和令牌预算 (~6K 令牌) 的双重限制
元数据跟踪 — 在
advisor-meta.jsonl中跟踪延迟、令牌估算和努力程度信号保护 — 被终止进程的部分输出会被丢弃,不会作为建议返回
路径遍历防护 — 文件读取经过验证,确保在项目根目录内
安全性
路径遍历保护:
files参数验证所有解析后的路径是否保留在项目根目录内。像../../etc/passwd或项目外的绝对路径将被拒绝。二进制文件过滤:常见的二进制扩展名(图像、可执行文件、存档等)会自动跳过。
无 Shell 执行:服务器使用带有数组参数的
spawn并通过 stdin 管道传输提示词。不会发生 Shell 插值。仅限本地:MCP 服务器通过 stdio 在本地运行。不会打开任何网络端口。
咨询日志:以纯文本形式存储在
~/.opus-advisor/中。这些文件可能包含您咨询中的代码片段和问题。如果这些文件包含敏感代码,请勿提交或共享它们。
环境变量
变量 | 描述 |
| 覆盖日志目录(默认: |
与 Anthropic 顾问工具的比较
Anthropic 的 advisor_20260301 是一项服务器端 API 功能,顾问可以在单个 API 请求中查看完整的对话记录。此 MCP 服务器采用不同的方法:
Anthropic 顾问工具 | opus-advisor-mcp | |
上下文共享 | 完整记录(服务器端) | 问题 + 文件 + 历史记录(客户端) |
身份验证 | 需要 API 密钥 | 使用现有的 Claude Code 订阅 |
集成 | API 级别 ( | MCP 工具(目前可在 Claude Code 中使用) |
持久性 | 无 | Markdown 日志 + JSONL 元数据 |
成本 | 按 Opus 费率的令牌计费 | 包含在订阅中 |
要求
Node.js >= 18
已安装并验证的 Claude Code CLI
许可证
MIT
Available Tools
4 toolsclear_advisor_logClear Advisor LogADestructive
Clear the consultation log and metadata to start fresh for this project.
Inputs: no named parameters.
consult_opusConsult Opus AdvisorARead-only
Consult Claude Opus 4.7 for strategic advice. Opus runs via the Claude Code CLI with your existing subscription — no API key needed. The advisor maintains a per-project consultation log for continuity across calls. History is capped by both entry count (5) and token budget (~6K tokens) to prevent context bloat. Use this for architecture decisions, complex debugging, code review, or any problem that benefits from deeper reasoning.
Inputs: question (required), context, effort, files, +1 more.
read_advisor_logRead Advisor LogARead-only
Read the consultation log from prior Opus advisor calls for this project. Useful for reviewing past advice or getting context on decisions already made.
Inputs: last_n.
read_advisor_metaRead Advisor MetadataARead-only
Read structured metadata (latency, token counts, effort levels) from all consultations. Useful for understanding cost and performance patterns.
Inputs: last_n.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 4 tools
Each tool has a uniquely defined purpose: clearing the log, consulting Opus, reading the log, and reading metadata. No two tools overlap in functionality, so an agent can easily distinguish them.
All tools follow a consistent verb_noun pattern using snake_case (e.g., consult_opus, read_advisor_log). The naming convention is uniform across the entire set.
Four tools is a well-scoped count for an advisor MCP server. Each tool addresses a distinct need (consultation, log management, metadata access) without unnecessary extras.
The tool set covers the core operations: consultation, log reading/clearing, and metadata inspection. A minor gap is the lack of a configuration tool to adjust Opus parameters, but the current surface is sufficient for most workflows.
Maintenance
Related MCP Connectors
Augments MCP Server - A comprehensive framework documentation provider for Claude Code
Shared memory MCP server for Claude Code, Codex, Gemini CLI and Cursor. Cited, editable.
Hosted X Ads MCP server for Claude, ChatGPT and Cursor, with each action switched per account.
Use AI models for chat, image, and video generation from Claude Code and other MCP hosts.
Related MCP Servers
- AlicenseAqualityBmaintenanceAn MCP server that lets Claude Code consult stronger AI models (o3, Gemini 2.5 Pro, DeepSeek Reasoner) when you need deeper analysis on complex problems.1148MIT
- AlicenseAqualityDmaintenanceMCP server that lets Claude Code ask GPT Codex for adversarial planning, code review, debugging, research, and risk triage without leaving your project workflow.910 npm2MIT
- AlicenseNot gradedqualityDmaintenanceAn MCP server that allows Claude Code to interact with the OpenAI Codex CLI.40 npm20MIT
- AlicenseNot gradedqualityBmaintenanceMCP server enabling Claude to consult Codex (GPT-5.x) mid-task for second opinions, plan/diff review, brainstorming, and codebase exploration via structured debates and permission-controlled interactions.2MIT