prompt-architect-mcp

Context-aware prompt enhancement for any LLM chat interface.

Remembers your project across conversations. Every prompt you send gets automatically enriched with the full history of what you have been building — no setup, no manual tracking. Works with OpenAI and Google Gemini.

How it works

Every time you call enhance_prompt_with_context, the server:

1. Saves your prompt to a local SQLite database
2. Reads everything you have discussed before in this project session
3. Builds a context block from your history and pinned facts
4. Enhances your prompt with that full context via OpenAI or Gemini
5. Saves the result — so the next call is even richer

You do one thing. The server handles everything else silently.

Requirements

• Node.js 18 or higher
• An OpenAI API key or a Google Gemini API key (or both)

Installation — Claude Desktop

Open your Claude Desktop config file:

• Windows: %APPDATA%\Claude\claude_desktop_config.json
• Mac: ~/Library/Application Support/Claude/claude_desktop_config.json

Using OpenAI only

{
  "mcpServers": {
    "prompt-architect": {
      "command": "npx",
      "args": ["-y", "prompt-architect-mcp@latest"],
      "env": {
        "OPENAI_API_KEY": "sk-your-openai-key"
      }
    }
  }
}

Using Gemini only

{
  "mcpServers": {
    "prompt-architect": {
      "command": "npx",
      "args": ["-y", "prompt-architect-mcp@latest"],
      "env": {
        "GEMINI_API_KEY": "your-gemini-key"
      }
    }
  }
}

Using both (OpenAI takes priority)

{
  "mcpServers": {
    "prompt-architect": {
      "command": "npx",
      "args": ["-y", "prompt-architect-mcp@latest"],
      "env": {
        "OPENAI_API_KEY": "sk-your-openai-key",
        "GEMINI_API_KEY": "your-gemini-key"
      }
    }
  }
}

Fully quit and reopen Claude Desktop after saving.

Installation — VS Code (GitHub Copilot Agent Mode)

Create or open:

• Global (all projects): %APPDATA%\Code\User\mcp.json (Windows) or ~/.config/Code/User/mcp.json (Mac/Linux)
• Workspace (this project only): .vscode/mcp.json in your project root

{
  "servers": {
    "prompt-architect": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "prompt-architect-mcp@latest"],
      "env": {
        "OPENAI_API_KEY": "sk-your-openai-key"
      }
    }
  }
}

Then open Copilot Chat (Ctrl+Alt+I), switch to Agent Mode, and your tools will appear.

Installation — Cursor

Create or open:

• Global: ~/.cursor/mcp.json
• Project: .cursor/mcp.json in your project root

{
  "mcpServers": {
    "prompt-architect": {
      "command": "npx",
      "args": ["-y", "prompt-architect-mcp@latest"],
      "env": {
        "OPENAI_API_KEY": "sk-your-openai-key"
      }
    }
  }
}

Or go to Cursor Settings → Tools & MCP → New MCP Server and paste the config.

Installation — Windsurf

Go to Windsurf Settings → MCP Servers → Add Server and paste:

{
  "mcpServers": {
    "prompt-architect": {
      "command": "npx",
      "args": ["-y", "prompt-architect-mcp@latest"],
      "env": {
        "OPENAI_API_KEY": "sk-your-openai-key"
      }
    }
  }
}

Installation — Firebase Studio (Project IDX / Antigravity)

Create .idx/mcp.json in your project root:

{
  "mcpServers": {
    "prompt-architect": {
      "command": "npx",
      "args": ["-y", "prompt-architect-mcp@latest"],
      "env": {
        "OPENAI_API_KEY": "sk-your-openai-key"
      }
    }
  }
}

Or use the Command Palette: Shift+Ctrl+P → Firebase Studio: Add MCP Server.

To keep your API key out of version control, put it in a .env file in your project root and omit the env block from mcp.json.

Tools

enhance_prompt_with_context — the main tool

Call this every time you want to enhance a prompt. Context grows automatically.

| Field | Required | Description | | ------------ | -------- | -------------------------------------------------------------------------- | | prompt | Yes | The raw prompt you want to enhance | | session_id | No | Your project name e.g. my-novel. Auto-detected from git root if omitted. |

First call — no context yet:

Session: my-api-project
No context yet — this is your first turn. Context will grow automatically from here.

Category: technical
Enhanced Prompt:
...

Second call onwards — context applied:

Session: my-api-project

Context applied:
[Project context]
History: User is building a Node.js REST API with Express and MongoDB...
Current task: Adding JWT authentication middleware
Pinned facts:
  - Using Node.js 20, MongoDB Atlas, RS256 algorithm

Category: technical
Enhanced Prompt:
...already knows your stack, your history, your decisions...

pin_fact — lock in a project decision

Use this when you want something remembered permanently across all future prompts in this project.

| Field | Required | Description | | ------------ | -------- | --------------------------------------------------------- | | fact | Yes | The fact to remember e.g. Using React 18 and TypeScript | | session_id | No | Must match the name used in enhance_prompt_with_context |

list_sessions — see all your projects

No fields required. Returns all active project sessions with their last active date.

enhance_prompt — quick one-off (no context)

For when you just want a single prompt enhanced without any session tracking.

| Field | Required | Description | | ------- | -------- | ------------------------- | | input | Yes | The raw prompt to enhance |

Testing the tools — step by step

Step 1 — Pin your project stack

Tool: pin_fact
fact: Building a task management app with Node.js 20, Express 4, MongoDB Atlas, JWT HS256
session_id: task-app

Expected: Fact pinned to session "task-app"

Step 2 — First enhancement (no history yet, but pinned fact appears)

Tool: enhance_prompt_with_context
prompt: create a mongoose schema for a task with priority levels
session_id: task-app

Expected: context block shows your pinned stack. Enhanced prompt mentions Node.js, MongoDB, Mongoose.

Step 3 — Second enhancement (context builds automatically)

Tool: enhance_prompt_with_context
prompt: write the Express route to create a new task
session_id: task-app

Expected: context block now shows history from Step 2. Enhanced prompt knows you already have a Mongoose schema.

Step 4 — Third enhancement (richer context)

Tool: enhance_prompt_with_context
prompt: add JWT middleware to protect the task routes
session_id: task-app

Expected: enhanced prompt references your schema and your routes from Steps 2 and 3.

Step 5 — Verify sessions

Tool: list_sessions

Expected:

Your project sessions:

• task-app   (last active: today)

Session ID guide

The session_id is just a name for your project. Rules:

• Use the same name every time for the same project
• Can contain letters, numbers, spaces, hyphens
• Spaces and special characters are converted to hyphens automatically
• If you are in a git repo and omit session_id, it is auto-detected from the repo name
• If not in a git repo and session_id is omitted, the tool will ask you to provide one

Good examples: my-novel, work-api-2025, recipe-app, thesis-project

Data location

All session data is stored locally on your machine at:

• Windows: C:\Users\YourName\.prompt-architect\sessions.db
• Mac/Linux: ~/.prompt-architect/sessions.db

No data is sent to any server other than your chosen LLM provider (OpenAI or Google) for prompt enhancement and summarisation.

License

MIT