
Cortex MCP — Multi-Level Reasoning Server
io.github.j0hanz由 GitHub 账号 j0hanz 发布v1.11.0更新于 Oct 24, 2026
Multi-level reasoning MCP server with configurable depth levels
概览
本地 stdio 服务器,为助手提供有状态、可配置深度的多步推理工具,并保留可查看的会话轨迹。
- 可以做什么
- 以 basic、normal、high 或 expert 深度运行并延续有步数上限的推理会话,可逐步进行或批量完成
- 按步骤记录结构化的观察、假设与评估,回滚到较早步骤,或提前给出结论结束会话
- 以资源形式读取活动会话列表、单个会话的 JSON 视图、完整 markdown 轨迹和单条思考
- 使用内置提示按指定深度开始推理、继续会话、重试失败任务或获取帮助
- 适用场景
- 适合希望助手以明确、有上限的步骤推演难题,并留下可查看或审计轨迹,而不是一次性给出推理的场景。
- 运行要求
- 需要 Node.js 24 或更高版本,通过 npx 在本地运行该 npm 包,以及支持 stdio 传输的 MCP 客户端。Docker 仅用于容器镜像,可选。无需账号、API 密钥或请求头;可选环境变量可调整会话有效期、会话数量、令牌上限和轨迹脱敏。
安装
在 SourceWeft 中
- 打开 控制台中的 Cortex MCP — Multi-Level Reasoning Server,将其添加到工作区。
- 为需要使用其工具的对话启用该服务。
Desktop only,通过 STDIO。 STDIO 服务会启动本地进程,因此需要 SourceWeft 桌面宿主。
其他 MCP 客户端
参照 仓库 中的启动说明。
README
Cortex MCP
[Install in VS Code] [Install in VS Code Insiders] [Install in Visual Studio]
[Add to LM Studio] [Install in Cursor] [Install in Goose]
Multi-level reasoning MCP server with configurable depth levels, session-based state management, structured thought input, and real-time trace resources.
Overview
Cortex MCP is a stdio-only MCP server for stateful, depth-controlled reasoning. The runtime entrypoint in src/index.ts connects createServer() to StdioServerTransport, and the server surface in src/server.ts enables tools, prompts, completions, logging, and subscribable resources around a single session-based reasoning engine.
The live MCP surface confirmed by Inspector is 1 tool, 6 concrete resources, 4 resource templates, and 7 prompts. Sessions are stored in memory, exposed as MCP resources, and cleared on process restart.
Key Features
reasoning_thinksupports step-by-step sessions,run_to_completionbatches, rollback, early conclusion, and structuredobservation/hypothesis/evaluationinput.- Four depth levels are built into the engine:
basic,normal,high, andexpert, each with bounded thought ranges and token budgets. - Prompt helpers expose
reasoning.basic,reasoning.normal,reasoning.high,reasoning.expert,reasoning.continue,reasoning.retry, andget-help. - Resource endpoints expose internal docs plus live session lists, per-session JSON views, full markdown traces, and individual thought documents.
- Completions are wired for levels, session IDs, and thought names through
completable()and resource-template completion hooks.
Requirements
- Node.js
>=24for localnpxornpmusage. - An MCP client that supports
stdiotransport. - Optional: Docker if you want to build or run the container image defined by
Dockerfile.
Quick Start
Use this standard MCP client configuration:
Client Configuration
Install in VS Code
Add to .vscode/mcp.json:
Or install via CLI:
For more info, see VS Code MCP docs.
Install in VS Code Insiders
Add to .vscode/mcp.json:
Or install via CLI:
For more info, see VS Code Insiders MCP docs.
Install in Visual Studio
Add to mcp.json (VS integrated):
For more info, see Visual Studio MCP docs.
Install in Claude Desktop
Add to claude_desktop_config.json:
For more info, see Claude Desktop MCP docs.
Install in Claude Code
Add to Claude Code CLI:
Or install via CLI:
For more info, see Claude Code MCP docs.
Install in Windsurf
Add to ~/.codeium/windsurf/mcp_config.json:
For more info, see Windsurf MCP docs.
Install in Codex CLI
Add to ~/.codex/config.yaml or codex CLI:
For more info, see Codex CLI MCP docs.
Install in Augment
Add to VS Code settings.json:
Add to your VS Code
settings.jsonunderaugment.advanced.
For more info, see Augment MCP docs.
Use Cases
Start bounded reasoning at the right depth
Use reasoning.basic, reasoning.normal, reasoning.high, or reasoning.expert when the client wants a prompt-first entrypoint, or call reasoning_think directly with query, level, and the first thought. Each response returns the current session state plus a summary string that tells the client how to continue.
Relevant tool: reasoning_think
Related prompts: reasoning.basic, reasoning.normal, reasoning.high, reasoning.expert
Continue, retry, or batch an active session
Reuse sessionId to continue a prior trace, switch to runMode="run_to_completion" when you already have the remaining thought inputs, or use the continuation and retry prompts to generate the next call payload. The handler also supports rollbackToStep and isConclusion for revising or ending a trace early.
Relevant tool: reasoning_think
Related prompts: reasoning.continue, reasoning.retry
Inspect live traces without re-running the tool
Read reasoning://sessions for the active session list, reasoning://sessions/{sessionId} for the JSON detail view, reasoning://sessions/{sessionId}/trace for the markdown transcript, or reasoning://sessions/{sessionId}/thoughts/{thoughtName} for a single thought. This lets a client present progress or audit a session independently from the next tool call.
Relevant resources: reasoning://sessions, reasoning://sessions/{sessionId}, reasoning://sessions/{sessionId}/trace, reasoning://sessions/{sessionId}/thoughts/{thoughtName}
Architecture
Request Lifecycle
MCP Surface
Tools
reasoning_think
Stateful reasoning tool for creating and continuing multi-step sessions. It supports one-step interactive calls, run_to_completion batches, structured observation/hypothesis/evaluation input, rollback, and early conclusion while returning structured session state.
Data Flow
Resources
Prompts
MCP Capabilities
Tool Annotations
Structured Output
reasoning_thinkdeclaresoutputSchemaand returnsstructuredContent, with an embedded trace resource when the trace is small enough. Evidence:src/tools/reasoning-think.ts:498,src/lib/mcp.ts:97-114.
Configuration
[!NOTE] The source does not define any HTTP host/port configuration. The only other environment-related signal is
NODE_ENV=productionin the Docker image and--env-file=.envin the localdev:runscript.
Security
[!NOTE] No auth, OAuth, HTTP origin checks, or rate-limiting controls are implemented in the current source because the server only exposes stdio transport.
Development
Additional helper scripts for diagnostics, coverage, asset copying, and knip are defined in package.json.
Build and Release
.github/workflows/release.ymlbumpspackage.jsonandserver.json, then runsnpm run lint,npm run type-check,npm run test, andnpm run buildbefore tagging and creating a GitHub release.- The same workflow publishes the package to npm with Trusted Publishing, publishes to the MCP Registry with
mcp-publisher, and pushes a multi-arch Docker image toghcr.io. Dockerfileuses a multi-stage Node 24 Alpine build, prunes dev dependencies, and runs the released container as themcpuser.
Troubleshooting
- Sessions are in memory and expire after 30 minutes by default. If you receive
E_SESSION_NOT_FOUND, start a new session or increaseCORTEX_SESSION_TTL_MS. runMode="run_to_completion"requires enoughthoughtentries to cover the remaining steps. If you want the server to return after each step, keep the defaultstepmode.- For stdio transport, do not add custom stdout logging around the server process. This server routes logs through MCP logging and falls back to
stderron failures.
Credits
Contributing and License
- License: MIT
- Contributions are welcome via pull requests.
来源:README.md,提交 c1171fd
工具
0版本历史
1- v1.11.0最新Sep 16, 2026
评分
已满足适用检查项的 40%
该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。
- 近 30 天内可连通25 分
请保持端点能够响应 MCP 握手;SourceWeft 每 30 天探测一次。
- 已声明认证方式10 分
请在注册中心条目中声明凭据请求头,或保持端点能够响应探测,以便判定其认证方式。
- 已列出工具15 分
请让服务器在无凭据时响应 tools/list,或发布可在沙箱中无凭据启动的软件包。
- 描述必需15 分
- 图标10 分
- 文档15 分
- 许可证10 分
请在仓库或注册中心条目中声明许可证。