ThreadMind MCP

Branches for your AI's memory: a git-friendly tree of thread summaries.

ThreadMind is a Model Context Protocol (MCP) server that gives AI coding sessions a memory: a tree of threads, each holding a short summary of what was decided about one topic, stored as Markdown in your repository. Start a fresh session and the active thread's context — its decisions and those of its ancestors — replaces a history that had grown to tens of thousands of tokens.

Documentation | npm | GitHub

Why ThreadMind?

With AI coding assistants, conversations grow long: every request carries the whole history, until the client compacts it or you start over — and starting over loses what was decided. ThreadMind keeps what matters:

• Summaries in sections — Decisions, Constraints, State, Open questions, Next steps
• Inheritance through the tree — a thread's context is its summary plus its ancestors' decisions and constraints; sibling branches stay out
• Branching exploration — try approaches in separate threads, merge the winner's conclusions into the parent, keep the reason an approach was abandoned
• Git-aware — a thread created on a feature branch becomes active whenever that branch is checked out, and summaries warn when the code they describe has changed
• Team collaboration — thread files are committed with the code and reviewed in pull requests

Where the savings come from

ThreadMind can't shrink a conversation in progress: your client decides what it sends. The savings come when you start over:

Before /clear:   the conversation carries ~48,000 tokens of history
After /clear:    the session starts with the thread's context, ~1,100 tokens

main
├── auth
│   └── auth-ui ← active     context = main + auth (decisions, constraints) + auth-ui
└── dashboard                 dashboard stays out

With the Claude Code plugin, the context reloads automatically after /clear and after compaction, and stats_show measures both sizes on your own sessions.

Quick Start

Claude Code plugin (recommended)

/plugin marketplace add mahmoud-nb/thread-mind-mcp
/plugin install thread-mind@thread-mind

The plugin installs the server and adds:

• a SessionStart hook that loads the active thread's context at startup, after /clear and after compaction (nothing happens in repositories without a ThreadMind project)
• a SessionEnd hook that records how large the conversation grew, read from its transcript, for stats_show
• commands: /thread-mind:context, /thread-mind:tree, /thread-mind:create, /thread-mind:switch, /thread-mind:summary, /thread-mind:merge, /thread-mind:stats

If you configured ThreadMind manually before, remove that configuration so Claude Code doesn't run two servers.

Installation

No installation required — run directly with npx:

npx thread-mind-mcp

Or install globally:

npm install -g thread-mind-mcp

Configure with Claude Code

Add to your Claude Code MCP settings (~/.claude/settings.json or project .claude/settings.json):

macOS / Linux:

{
  "mcpServers": {
    "thread-mind": {
      "command": "npx",
      "args": ["-y", "thread-mind-mcp"]
    }
  }
}

Windows:

{
  "mcpServers": {
    "thread-mind": {
      "type": "stdio",
      "command": "cmd",
      "args": ["/c", "npx", "thread-mind-mcp"],
      "env": {}
    }
  }
}

On Windows, npx must be wrapped with cmd /c because npx is a .cmd wrapper and cannot be spawned directly by the MCP stdio transport.

Windows + Volta:

If you use Volta as your Node.js version manager, use volta run to ensure the correct Node.js version is resolved when Claude Code spawns the MCP subprocess:

{
  "mcpServers": {
    "thread-mind": {
      "type": "stdio",
      "command": "cmd",
      "args": ["/c", "volta", "run", "npx", "-y", "thread-mind-mcp"],
      "env": {}
    }
  }
}

Or via CLI: claude mcp add thread-mind-mcp --scope project -- cmd /c volta run npx -y thread-mind-mcp

Configure with other MCP clients

ThreadMind uses the stdio transport, compatible with any MCP client. Use the same configuration above for your platform.

Workspace location

ThreadMind stores its data in .threadmind/ at the root of your workspace, found in this order:

1. The THREADMIND_ROOT environment variable, if set
2. The workspace roots advertised by the MCP client (a root that already contains .threadmind/ wins)
3. The directory the server was started from

A client that starts servers outside your project without advertising roots (for example a desktop chat app with no project folder) needs THREADMIND_ROOT:

{
  "mcpServers": {
    "thread-mind": {
      "command": "npx",
      "args": ["-y", "thread-mind-mcp"],
      "env": { "THREADMIND_ROOT": "/path/to/your/project" }
    }
  }
}

Teach your AI client to use ThreadMind

The server sends its usage instructions when the client connects (MCP server instructions): Claude Code needs no setup file. For agents that don't read them, run threadmind_init to generate an AGENTS.md file, read by Codex, Cursor, GitHub Copilot and others.

How It Works

Core Concepts

Storage

ThreadMind stores everything in a .threadmind/ directory at your project root:

.threadmind/
  config.json              # Local state: where sessions start, author ID, session measurements
  .gitignore               # Excludes config.json from git
  projects/
    my-app.json            # Project configuration
  threads/
    my-app/
      main.md              # Root thread (markdown + YAML frontmatter)
      auth-system.md       # Child thread
      auth-api.md          # Grandchild thread

Thread files use YAML frontmatter. Each one records its parent: the tree is rebuilt from the files, with no index to keep in sync or to conflict in git.

---
id: auth-system
title: Authentication System
parentId: main
author: mahmoud-a3f9
createdAt: 2026-04-15T10:00:00.000Z
updatedAt: 2026-04-15T12:30:00.000Z
branch: feature/auth
commit: 3f2a9c1e5b7d4a8f0c6e2b1d9a7f5c3e1b0d8a6f
paths: ["src/auth"]
---

## Decisions
- JWT with refresh tokens, stored in httpOnly cookies
- Passport.js over custom middleware, for maintainability

## State
- Login and refresh done, logout pending

status, branch, commit, paths and mergedInto only appear when set.

Context Assembly

When you request context, ThreadMind walks up from the active thread to the root. It keeps the whole summary of the active thread, and only the Decisions and Constraints of its ancestors:

## System Context
You are building a Next.js e-commerce application...

---

## Thread: My App

## Decisions
- Next.js 15, PostgreSQL, Stripe

---

## Thread: Authentication System

## Decisions
- JWT with refresh tokens, bcrypt, Passport.js

---

## Thread: Auth API Endpoints (active)

_⚠ 2 commits changed src/api since this summary was written: check it is still accurate._

## Decisions
- POST /auth/login, POST /auth/register, POST /auth/refresh

## Next steps
- Rate limiting on /auth/login

---
_ThreadMind context: ~160 tokens | depth: 3 threads_

• Only the direct ancestor chain is included — sibling branches are excluded
• Summaries written before code changes are flagged, using the commit and paths recorded with them
• maxTokens leaves out the farthest ancestors first when the context would be too large

Available Tools

Each tool declares a title and MCP annotations (read-only, destructive, idempotent), which clients can use to decide when to ask for confirmation.

Project Management

project_create

Thread Management

thread_create

thread_status

thread_merge

thread_delete

thread_rebase

Summary & Context

summary_update

context_get

Setup

threadmind_init

Writes the same instructions the server sends to MCP clients:

Earlier versions wrote CLAUDE.md and .cursorrules. threadmind_init moves the ThreadMind section out of .cursorrules when it generates the Cursor rule, and reports any section left in either file.

Statistics

With the plugin, each session records the size its conversation reached (read from the Claude Code transcript) and the context loaded when the next one started. stats_show reports both:

Measured sessions (Claude Code plugin):
  Conversation size when sessions ended: ~48,200 tokens on average, ~96,000 at most (12 sessions)
  ThreadMind context loaded at session start: ~1,150 tokens on average (14 sessions)

Available Resources

Every thread of the active project is listed as a resource: in Claude Code, type @ to mention one and bring a sibling thread's summary into the conversation.

Available Prompts

tm-context, tm-tree and tm-stats embed their result directly, without a tool call. The other prompts ask the AI to call the matching tool, so your client's confirmation settings still apply to changes. Thread ID arguments support autocompletion.

In Claude Code, these appear as slash commands: /mcp__thread-mind__tm-help, /mcp__thread-mind__tm-create, etc.

Quick Shortcuts

The server instructions teach the AI short text commands you can type directly in chat:

Usage Examples

1. Start a new project

You: Create a new ThreadMind project called "E-Commerce App" with system context
     "Building a Next.js e-commerce platform with Stripe payments"

AI: [calls project_create] → Project "e-commerce-app" created. Main thread active.

Claude Code needs no further setup. To let other agents (Codex, Cursor, Copilot…) use the project too, ask for threadmind_init, which generates an AGENTS.md file.

2. Work and summarize

You: [discuss the architecture with AI...]
You: tm:summary

AI: [writes the summary in sections, then calls summary_update]
     ## Decisions
     - Next.js 15 App Router, PostgreSQL with Prisma, Stripe Checkout
     ## Next steps
     - Authentication

You: We'll use Stripe webhooks for order confirmation, record it.

AI: [calls summary_update with section "decisions"] → Section "decisions" updated.

3. Branch into a sub-topic, on its git branch

$ git switch -c feature/payments

You: tm:create Payment Integration

AI: [calls thread_create] → Thread "payment-integration" created under "main".
     Linked to git branch "feature/payments".

     main
     └── payment-integration [branch: feature/payments] ← active

4. Start fresh when the conversation grows

You: tm:summary
You: /clear

(the plugin loads the context of payment-integration: main's decisions
 and constraints, then the whole payment-integration summary)

Without the plugin, ask for tm:context in the new session.

5. Finish the thread

You: tm:merge

AI: [writes main's new summary with the payment decisions, calls thread_merge]
     → Thread "payment-integration" merged into "main", which is now the active thread.

     main ← active
     └── payment-integration [done] [branch: feature/payments]

Team Mode

Team mode enables collaborative thread trees shared via git.

How it works

1. Create a project in team mode:

   Code

   Copy

   project_create with title "Shared Project" and mode "team"
   

2. Each team member gets an author ID on their first write, derived from git config user.name and a short hash of git config user.email (stable across clones; override it with the THREADMIND_AUTHOR environment variable)

3. Thread files (.threadmind/threads/) are tracked by git. Each one records its parent, so two people creating threads never conflict, even under the same parent

4. The local config (.threadmind/config.json) is gitignored — each member has their own session state and measurements

5. A teammate who checks out your feature branch starts on the thread linked to it

Rules

Ownership prevents accidental edits; it is not access control.

Upgrading from 0.4: the tree used to live in .threadmind/trees/. ThreadMind now moves it into the thread files on first use and deletes trees/<project>.json — commit that change, and upgrade the whole team together. .threadmind/stats/ is no longer used and can be deleted.

Workflow

# Pull teammates' threads
git pull

# View the full tree (includes everyone's threads)
# → Use thread_list

# Branch from a teammate's thread
# → Use thread_create with parentId set to their thread

# Push your new threads
git add .threadmind/
git commit -m "Add payment-integration thread"
git push

Development & Publishing

Development setup, tests, releases and publishing to the MCP Registry are described in SETUP.md.

Architecture

src/
  index.ts              # Entry point — stdio server, or `hook <event>` for the Claude Code hooks
  server.ts             # McpServer factory (tools + resources + prompts + instructions)
  services.ts           # Services over one .threadmind directory, shared by server and hooks
  types/
    index.ts            # All TypeScript interfaces
  core/
    frontmatter.ts      # YAML frontmatter parser/serializer (zero deps)
    ids.ts              # ID validation, slugification, title normalization
    lock.ts             # Cross-process lock file
    clock.ts            # Strictly increasing timestamps
    storage.ts          # File I/O with atomic writes; tree derived from thread files
    workspace.ts        # Workspace root resolution (env, client roots, cwd)
    session.ts          # Active thread per session, per git branch
    git.ts              # Current branch, commits since a summary
    summary.ts          # Summary sections: inheritance, section updates
    project.ts          # Project lifecycle management
    thread.ts           # Thread CRUD, status, merge, tree rendering
    context.ts          # Context assembly, outdated summary notes, token budget
    instructions.ts     # Server instructions and instruction files
    stats.ts            # Session measurements and statistics report
  hooks/
    index.ts            # SessionStart / SessionEnd hooks of the Claude Code plugin
  tools/
    index.ts            # 14 MCP tools with Zod schemas and annotations
  resources/
    index.ts            # 2 MCP resources + the thread resource template
  prompts/
    index.ts            # 11 MCP prompts
plugin/                 # Claude Code plugin (manifest, .mcp.json, hooks, commands)
.claude-plugin/         # Marketplace listing the plugin
server.json             # MCP Registry metadata

Design Decisions

• File-based storage over SQLite — git-friendly, human-readable, zero native dependencies
• Thread files as the single source of truth — each file records its parent; no tree index to drift or to conflict in git
• YAML frontmatter — thread metadata and content in a single .md file, readable by both humans and tools
• No external YAML parser — minimal hand-rolled parser for the simple flat frontmatter format
• Atomic writes — write to a temp file, flush it, then rename it over the target, so a crash never leaves a partial file
• Serialized updates — read-modify-write operations run one at a time, across sessions and hooks too, through a lock file
• Slugified IDs — thread IDs derived from titles ("Auth System" → "auth-system"), collision-safe with auto-suffix, restricted to a-z, 0-9 and - (which also rules out path traversal)
• One active thread per session — sessions don't move each other; new sessions start from the checked-out git branch
• Selective inheritance — ancestors pass on their decisions and constraints, not their state or next steps
• MCP Prompts — slash commands; read-only ones embed their result instead of asking for a tool call
• Server instructions — the server tells clients how to use it, so no instruction file is needed; threadmind_init writes the same text to AGENTS.md and other files for agents that don't read them
• Workspace resolution — THREADMIND_ROOT, then the client's roots, then the working directory
• Measured, not estimated, savings — conversation sizes come from the Claude Code transcripts; context sizes are estimates (~3.5 characters per token)

Requirements

• Node.js >= 18.0.0
• Git (optional: team mode author detection, thread–branch links, outdated summary warnings)

License

MIT