kiro-cli --resume-id cli_3342dc00-47e5-456b-aaae-b62f0afcdcac_CvkJS8dD

MCP Context Manager

A local MCP server for AI session context management — transparent, offline, no black box.

Overview

mcp-context-manager is a local Model Context Protocol server written in TypeScript that helps AI assistants manage their context window during long sessions. It exposes 5 tools to measure, summarize, clean up, export and resume session context.

Features

Tool	Description
estimate_tokens	Estimates token usage and returns ok / warning / critical status
summarize_context	Extracts a structured Markdown summary from session text
suggest_cleanup	Identifies verbose or redundant content to remove
export_summary	Saves the summary to a timestamped Markdown file
load_session	Lists and loads previous session files for context resumption

All processing is local — no external LLM, no network calls, heuristics and regex only. Works completely offline.

---

Context thresholds

Based on field-tested experience:

Status	Threshold	Recommended action
ok	< 60%	No action needed
warning	60–90%	Run summarize_context + export_summary
critical	> 90%	Export immediately before context loss

---

Installation

Prerequisites

• Node.js 24+
• npm

Build from source

npm install
npm run build

---

Configuration

Kiro (.kiro/settings/mcp.json)

{
  "mcpServers": {
    "context-manager": {
      "command": "node",
      "args": ["./mcp-context-manager/dist/index.js"]
    }
  }
}

Claude Desktop (claude_desktop_config.json)

{
  "mcpServers": {
    "context-manager": {
      "command": "npx",
      "args": ["tsx", "./mcp-context-manager/src/index.ts"]
    }
  }
}

---

Usage

Estimate current context usage

Call estimate_tokens with your session text.
→ Returns: estimated_tokens, usage_percent, status, recommendation

Summarize and export before clearing

1. Call summarize_context with your session text
2. Review the summary — especially the "## Resume Here" section
3. Call export_summary with the summary and a subject
→ Returns: file path in ./context-summaries/
4. Ask the user to run /clear

Resume a previous session

1. Call load_session (no arguments) to list available sessions
2. Confirm which session to load with the user
3. Call load_session with the file name
→ Returns: session content ready to inject as context

---

Session files

Exported summaries are saved in ./context-summaries/ with the format:

session-YYYY-MM-DD-<subject>.md

Each file contains 6 mandatory sections:

## Main Objective
## Decisions Made
## Modified Files
## Next Steps
## Warnings
## Resume Here

The ## Resume Here section contains the suggested first prompt for the next session — the most critical section for context resumption.

⚠️ Security note: Session files may contain sensitive information discussed during the session (architecture decisions, credentials mentioned, etc.). Add context-summaries/ to your .gitignore to prevent accidental commits.

---

Project structure

mcp-context-manager/
├── src/
│   ├── index.ts              # MCP server entry point
│   ├── tools/
│   │   ├── tokens.ts         # estimate_tokens
│   │   ├── summarize.ts      # summarize_context
│   │   ├── cleanup.ts        # suggest_cleanup
│   │   ├── export.ts         # export_summary
│   │   └── load.ts           # load_session
│   └── utils/
│       └── tokenizer.ts      # Token estimation utility
├── plugin.json               # Kiro Power manifest
├── mcp.json                  # MCP server configuration
├── skills/
│   └── context-export/
│       └── SKILL.md          # Context export workflow
└── dev.kiro/
    └── steering/
        └── context-manager.md

---

Kiro Power

This project is also a Kiro Power — install it to get the context-export Skill and the MCP server loaded automatically when you mention keywords like context, tokens, session, export session, or resume session.

---

Development

Run in development mode

npm run dev

Run tests

npm test

Test with MCP Inspector

npx @modelcontextprotocol/inspector npx tsx src/index.ts

---