TodoMCP

TodoMCP

A local Model Context Protocol server that helps Codex turn requests into atomic tasks, manage plans, and rigorously verify completions without using any model API.

Category
访问服务器

README

TodoMCP

CI

TodoMCP is a local Model Context Protocol server that helps Codex plan genuinely complex work and verify uncertain completion claims without burdening small, deterministic tasks with planning overhead.

It uses no model API, opens no port, and has no dependency on another MCP server. State is stored outside your repositories in the operating system's user-data directory.

TodoMCP complements CountdownMCP, but both servers remain fully independent and useful on their own.

What it does

  • Creates stable source units from numbered lists, bullets, sentences, compound clauses, constraints, and success language.
  • Leaves small, clear, deterministic tasks entirely to Codex with zero TodoMCP calls.
  • Audits uncertain small results with one stateless call and no stored plan or task.
  • Requires a validated atomic plan for larger, dependent, risky, or multi-system work.
  • Maps every source unit to an explicit requirement disposition.
  • Blocks tasks whose dependencies are incomplete.
  • Audits the verification method as well as its result.
  • Recommends zero to three bounded agent work packages without creating agents.
  • Exports the neutral WorkCandidate.v1 shape for optional CountdownMCP advice.

Requirements

  • Node.js 20 or newer
  • Codex CLI/Desktop with MCP support

Install

Clone the repository and run the installer:

git clone https://github.com/AlonTsur1601/TodoMCP.git
Set-Location TodoMCP
node install.mjs

The installer builds and smoke-tests the server before atomically copying it to CODEX_HOME/mcp/todo-mcp. It replaces only the todo_mcp registration and rolls back that registration and the installed files if setup fails. Restart Codex after installation.

An installed copy is a local snapshot. New Git commits and GitHub Releases do not update it automatically. To update an existing clone, pull the newer source and rerun the installer:

Set-Location TodoMCP
git pull --ff-only
node install.mjs

Restart Codex after installing or updating.

For development without installing:

npm ci
npm run check
codex mcp add todo_mcp -- node C:\absolute\path\to\TodoMCP\dist\src\index.js

Workflow

  1. For an obviously small deterministic task, Codex works independently and makes no TodoMCP calls.
  2. For small work with meaningful completion uncertainty, Codex works independently and may call todo_audit_result once at the end. This creates no persistent TodoMCP state.
  3. For genuinely complex work, todo_analyze_request returns stable source units and recommends a plan.
  4. Codex maps every unit to a requirement and submits todo_create_plan once.
  5. todo_audit_completion checks a planned task and can auto-start it when its dependencies are ready, avoiding a separate start call.
  6. todo_close_plan is called once after the final approved task.

Rejected plans are retained as drafts so Codex can repair them with todo_revise_plan. There is no force-complete operation.

Small tasks and tool-call budget

Small does not mean that Codex must create a shorter TodoMCP plan. It means Codex should work without TodoMCP state. A precise edit with an immediately observable result, such as replacing one known line with exact text, needs no TodoMCP analysis, task creation, start, audit, or close call.

If a small task has meaningful uncertainty—such as runtime behavior, a bug fix, an integration boundary, or user-visible behavior—Codex may make one todo_audit_result call after doing the work. The caller must state why completion is uncertain. If neither the request nor the caller supplies an uncertainty reason, the tool skips the audit.

The legacy direct plan input remains accepted for backward compatibility, but clients should not create direct plans for ordinary small work.

Verification quality

For planned work and uncertain independent results, each acceptance criterion declares its minimum evidence level: static, build, unit, integration, runtime, or manual. Evidence must include the target, input, expected and observed signals, substantive raw output, and an explanation of how the check detects failure.

Behavioral criteria cannot be satisfied by build-only evidence. User-visible criteria can require observation through the public interface, and criteria can require negative or boundary cases. Failure-sensitivity explanations must identify a concrete failing, changed, rejected, negative, or pre-fix signal. Explicit artifact paths are confined to the registered workspace, limited to 1 MiB, and SHA-256 checked. Artifacts declared as test_source are also inspected for a recognizable assertion or failure check, so a script that only prints PASS is rejected.

TodoMCP cannot make an uncooperative MCP client call its tools or independently understand arbitrary verification code without a model. Its state machine and evidence contract provide the strongest enforcement available inside a model-free MCP tool workflow.

CountdownMCP interoperability

todo_get_execution_candidates returns:

{
  "schemaVersion": "WorkCandidate.v1",
  "currentTaskId": "task-2",
  "tasks": [
    {
      "id": "task-2",
      "title": "Add runtime verification",
      "priority": 5,
      "estimatedMinutes": 45,
      "dependenciesReady": true,
      "needsUserInput": false,
      "canContinueWithoutNewMessage": true,
      "checkpointable": true
    }
  ]
}

Codex may pass this result to countdown_advise_work, then pass the returned recommendedNow, deferUntilReset, executionOrder, and optional checkpoint to todo_apply_execution_advice. Advice can reorder ready work but cannot bypass dependencies or verification. Either server works normally when the other is absent.

Data and security

  • Windows: %LOCALAPPDATA%\TodoMCP\workspaces
  • macOS: ~/Library/Application Support/TodoMCP/workspaces
  • Linux: $XDG_DATA_HOME/todo-mcp/workspaces or ~/.local/share/todo-mcp/workspaces

Tool input is untrusted. TodoMCP uses strict schemas, atomic state writes, per-plan locks, path containment checks, and stderr-only diagnostics. It does not execute verification commands or read another MCP server's state.

Development

npm run typecheck
npm test
npm run smoke

Live account-dependent tests are intentionally unnecessary: TodoMCP has no account or OpenAI API integration.

When a sibling CountdownMCP checkout has been built, verify the real tool contract without adding a package dependency:

npm run contract:countdown

Releases

Pushing a tag that exactly matches the version in package.json (for example, v0.1.0) runs the release workflow. It repeats the full check and publishes a GitHub Release containing a source ZIP and SHA-256 checksum. Users extract the ZIP and run node install.mjs; no global npm publication is required.

License

MIT License

推荐服务器

Baidu Map

Baidu Map

百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。

官方
精选
JavaScript
Playwright MCP Server

Playwright MCP Server

一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。

官方
精选
TypeScript
Magic Component Platform (MCP)

Magic Component Platform (MCP)

一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。

官方
精选
本地
TypeScript
Audiense Insights MCP Server

Audiense Insights MCP Server

通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。

官方
精选
本地
TypeScript
VeyraX

VeyraX

一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。

官方
精选
本地
graphlit-mcp-server

graphlit-mcp-server

模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。

官方
精选
TypeScript
Kagi MCP Server

Kagi MCP Server

一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。

官方
精选
Python
e2b-mcp-server

e2b-mcp-server

使用 MCP 通过 e2b 运行代码。

官方
精选
Neon MCP Server

Neon MCP Server

用于与 Neon 管理 API 和数据库交互的 MCP 服务器

官方
精选
Exa MCP Server

Exa MCP Server

模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。

官方
精选