@anthropic-claude/later

MCP Server for Deferred Decision Management

Capture decisions with full context, revisit later without memory overhead. A production-ready Model Context Protocol (MCP) server following the MCP 2025-06 specification.

The Problem

During complex work, you encounter decisions that:

• Aren't urgent enough to do now
• Require significant context to understand later
• Get lost in conversation history
• Are expensive to reconstruct from memory

Example: "Should I optimize CLAUDE.md (3,343 words) now or later? What were the options? What was recommended?"

The Solution

# Capture decision with context
later_capture: "Optimize CLAUDE.md size"

# Later... review deferred items
later_list

# Get full context
later_show 1

# Start working on it
later_do 1

Installation

npm (Global)

npm install -g @anthropic-claude/later

Claude Code Integration

Add to your Claude Code MCP configuration (~/.claude/.mcp.json):

{
  "mcpServers": {
    "later": {
      "command": "npx",
      "args": ["@anthropic-claude/later"],
      "env": {}
    }
  }
}

Or if installed globally:

{
  "mcpServers": {
    "later": {
      "command": "later-mcp",
      "env": {}
    }
  }
}

From Source

git clone https://github.com/chudeemeke/later.git
cd later
npm install
npm run build

Features

Progressive Tool Disclosure

Only the search_tools meta-tool is exposed initially, reducing token overhead by ~90%. Other tools are discovered on-demand:

search_tools: "create a decision"
# Returns: later_capture tool with schema

Automatic PII Protection

Sensitive data in context is automatically detected and tokenized:

• API keys (OpenAI, Anthropic, GitHub, etc.)
• Passwords and secrets
• Social Security Numbers
• Credit card numbers
• Email addresses
• IP addresses
• Phone numbers

Detection rate: 95%+ accuracy with zero false negatives on tested patterns.

Available Tools

| Tool | Description | |------|-------------| | search_tools | Discover tools based on what you want to do | | later_capture | Capture a decision with context | | later_list | List and filter deferred items | | later_show | Show full item details | | later_do | Mark item as in-progress, get todo guidance | | later_update | Update any item field | | later_delete | Soft or hard delete items | | later_bulk_update | Update multiple items at once | | later_bulk_delete | Delete multiple items at once | | later_search | Full-text search with relevance ranking |

MCP 2025-06 Compliance

• outputSchema for all tools (typed responses)
• Structured error handling with isError pattern
• All logging to stderr (stdout reserved for protocol)
• Graceful shutdown handling
• Progressive tool disclosure

CLI Usage

# Capture a decision
later capture "Refactor authentication module"

# Capture with context and priority
later capture "Migrate to PostgreSQL" --context "Current MySQL has scaling issues" --priority high --tags "database,migration"

# List all items
later list

# Filter items
later list --status pending --priority high
later list --tags "refactor"

# Show item details
later show 1

# Start working on an item
later do 1

# Update item
later update 1 --status done
later update 1 --priority medium --tags "completed,v2"

# Delete (soft - archives)
later delete 1

# Delete (hard - permanent)
later delete 1 --hard

# Search
later search "authentication"

Data Storage

Items are stored in ~/.later/items.jsonl:

• JSONL format (one JSON object per line)
• Atomic writes with file locking
• Automatic corruption recovery
• Secure file permissions (600)

API Reference

later_capture

{
  decision: string;      // Required: Decision text (1-500 chars)
  context?: string;      // Optional: Additional context
  tags?: string[];       // Optional: Categorization tags
  priority?: 'low' | 'medium' | 'high';  // Default: medium
}

later_list

{
  status?: 'pending' | 'in-progress' | 'done' | 'archived';
  tags?: string[];       // Filter by tags (OR logic)
  priority?: 'low' | 'medium' | 'high';
  limit?: number;        // Default: 50
}

later_search

{
  query: string;         // Required: Search query
  status?: string;       // Filter by status
  tags?: string[];       // Filter by tags
  priority?: string;     // Filter by priority
  limit?: number;        // Default: 10
}

Development

# Install dependencies
npm install

# Build
npm run build

# Run tests
npm test

# Run tests with coverage
npm run test:coverage

# Lint
npm run lint

# Watch mode
npm run watch

Test Coverage

• Statements: 95.91%
• Branches: 90.26%
• Functions: 97.76%
• Lines: 95.95%

All coverage metrics exceed industry standards (80-85% for branches).

Architecture

src/
  index.ts              # MCP server entry point
  registry.ts           # Tool registry for progressive disclosure
  types.ts              # TypeScript type definitions
  schemas/              # Output schemas (MCP 2025-06)
  storage/              # JSONL storage with file locking
  tools/
    core/               # capture, list, show
    workflow/           # do, update, delete
    batch/              # bulk operations
    search/             # full-text search
    meta/               # search_tools
  utils/                # Utilities (logger, validation, etc.)
  cli/                  # CLI client

Security

• Automatic PII detection and tokenization
• Secret sanitization (API keys, passwords)
• Secure file permissions (600/700)
• Input validation with Zod schemas
• No SQL injection risk (JSONL storage)

Requirements

• Node.js >= 18.0.0
• Compatible with Claude Code, Claude Desktop, or any MCP client

License

MIT

Contributing

Contributions welcome! Please read our contributing guidelines and submit PRs to the main branch.

Changelog

See CHANGELOG.md for release history.

Support

• Issues: https://github.com/chudeemeke/later/issues
• Documentation: https://github.com/chudeemeke/later#readme

Built with the Model Context Protocol SDK.