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.
README
TodoMCP
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.v1shape 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
- For an obviously small deterministic task, Codex works independently and makes no TodoMCP calls.
- For small work with meaningful completion uncertainty, Codex works independently and may call
todo_audit_resultonce at the end. This creates no persistent TodoMCP state. - For genuinely complex work,
todo_analyze_requestreturns stable source units and recommends a plan. - Codex maps every unit to a requirement and submits
todo_create_planonce. todo_audit_completionchecks a planned task and can auto-start it when its dependencies are ready, avoiding a separate start call.todo_close_planis 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/workspacesor~/.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
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.