rubber-duck-mcp

I am your persistent, project-aware coding assistant. It helps developers remember, automate, and orchestrate project tasks. (eventually, ha)

think of it as a smart rubber duck for your code — it listens, remembers, and helps you get things done efficiently.

Running rubber-duck-mcp locally

Pre-requisites

1. Install and set up nvm - optional
2. Install node version v24.13.1. nvm install v24.13.1
3. you will need python2 to compile better-sqlite3 package. if you have pyenv, you can set both python2 and python3 version

Steps to run server

1. clone repo git@github.com:kencanak/rubber-duck-mcp.git
2. boot up the server by running ./run.sh in terminal, this tool is running on node 24 and using nvm to set the node version.

Using rubber-duck mcp in your project

1. rename .mcp.json.sample > .mcp.json, then copy it to your respective project folder. which ever project that you want to use this mcp server with
2. copy claude.md from .claude folder to the respective project folder.
3. for some reason, claude doesn't always read the claude.md file. so, we might need to nudge it.
4. video ref

Available tools

1. duck-recall

A project-specific, persistent developer memory system. It allows developers and AI assistants like Claude to store, recall, and manage important project knowledge across sessions. Think of it as a “Rubber Duck” that remembers key decisions, conventions, debugging discoveries, and tooling preferences for your project — so you don’t have to repeat yourself.

a. why?

• Captures institutional knowledge: Store decisions about architecture, tooling, or coding conventions.
• Speeds up onboarding: New team members or assistants can recall previous decisions instantly. (THOUGH, it's still stored in local db - to be evolved)
• Reduces repetition: Avoid re-explaining recurring problems or solutions.
• Supports AI-assisted development: AI agents can recall relevant memories to provide smarter suggestions and context-aware guidance.

b. how?

• store memory: You or an AI can store a memory with a summary, category, confidence, and optionally files to reference.
• Recall Memory: Memories can be retrieved for the current project, sorted by confidence and recent usage.
• Search Memory: Memories can be search by query. this way, we avoid using all search result as token
• Reinforce or Contradict: Update confidence based on confirmations or corrections, keeping the memory relevant.
• Archive: Retire outdated or irrelevant memories.
• Everything is scoped per project, so memories are automatically relevant to the current working directory without needing global identifiers. This allows AI assistants to provide context-aware responses and maintain a persistent understanding of project-specific knowledge.

c. with Claude:

we set a base prompt. available in .claude/claude.md

d. motivation behind this:

• instead of writing memory into a markdown file, i find it cleaner if we store it into DB
• which then can be evolved to vectorised db, a vectorized DB stores embedding vectors of memory content and allows semantic similarity search rather than literal text match. but of course, this comes with complexity
• hoping that we can use less token? ha

2. TODO:

add a couple more tools:

1. parallel development orchestration hub - where each “agent” is a lightweight process acting on a separate workspace (Git worktree) to perform parallel tasks.

• Each agent gets its own Git worktree (or branch) to avoid conflicts.
• MCP orchestrates agents with a task queue:
  • lint & fix
  • feature development
  • bug investigation
  • test & deploy
• Agents can read project memory from Duck-Recall and update it if needed.
• Results are PRs

2. create a workflow runner?
3. task scheduler?

• daily code analysis: run linters, tests, and code complexity reports every night.
• memory cleanup: archive or reinforce old duck-recall memories periodically.
• auto-deploy branches: on commit to a staging branch, trigger workflow that builds and deploys.