MDPilot

Markdown intelligence platform — generate production-grade AI instruction files, optimize tokens, and convert any file to markdown. Works in the browser and as an MCP server inside your IDE.

mdpilot.in · Docs · MCP setup · VS Code extension · npm — mdpilot-mcp@0.3.3 · Changelog

---

What it generates

Mode	Input	Output
Task (primary)	Ticket · Slack thread · spec	TASK.md + SPEC.md — structured, gap-checked, agent-ready
Generate	7-step wizard + tech stack	README · AGENTS · CLAUDE · CONTRIBUTING · SECURITY · SKILL · DESIGN · CONTEXT
Convert	PDF · DOCX · CSV · HTML	Clean markdown via MarkItDown
Explain	Code file or directory	WALKTHROUGH.md tuned to any audience
Image → Prompt	Any image	Recreation prompt for FLUX / SD / Midjourney / DALL-E
Interview primer	Role + level + JD	Ready-to-paste AI coach prompt

Every output runs through a 5-pass token optimizer (20–40% reduction) and a self-verification loop that checks generated commands against your real repo before writing anything.

DESIGN.md follows the getdesign.md / awesome-design-md ecosystem spec — 9 sections in fixed order (Visual Theme → Agent Prompt Guide) so any design tool or AI agent that reads DESIGN.md files gets exact tokens, not guessed values.

Agent behavior directives (v0.3.1):

• Human voice (writingStyle: 'human') — natural prose for README, CONTRIBUTING, DESIGN; auto-enabled for non-technical and learner audiences; hard-ignored for agent files (AGENTS.md, CLAUDE.md, SKILL.md).
• Plan-first (ai_exec mode) — generated agent prompts now always open with "Before executing: write a 3-5 line plan" + step-verify gates before the Response style line.
• Agent risk check (ai_exec + riskCheck toggle) — appends "check your plan against the Watch-outs above" to the agent prompt block when Watch-outs are present.

---

VS Code extension — no config required

Install the MDPilot extension from the VS Code Marketplace. Works in VS Code, Cursor, and Windsurf. No JSON MCP config — just install, and the extension auto-reads your existing API key from ~/.cursor/mcp.json, ~/.claude/mcp.json, or env vars.

A sidebar panel gives you a Chat tab (type generate agents, check drift, task: …) and a Settings tab (provider, encrypted key, status) — or use the command palette:

Ctrl+Shift+P → MDPilot: Generate AGENTS.md

Command	What it does
MDPilot: Generate AGENTS.md	Verified AGENTS.md, written to disk
MDPilot: Generate CLAUDE.md	CLAUDE.md for Claude Code / Cursor
MDPilot: Generate Task Prompt	Ticket → TASK.md
MDPilot: Check Docs for Drift	Flags stale commands and paths
MDPilot: Save / Load Session Context	Persist session state across AI sessions
MDPilot: Open Settings	Provider, key, and connection status in the panel

Extension setup → · Marketplace →

---

MCP server — use MDPilot from inside your IDE

The mdpilot-mcp npm package exposes all 10 tools directly to Claude Code, Cursor, Windsurf, and Goose. The server reads your actual repo on disk — generated files reference real scripts and paths, never guesses.

One-command setup (gets you a free key + writes the config):

npx -y mdpilot-mcp setup

Or non-interactively:

npx -y mdpilot-mcp setup --key gsk_... --client claude

Manual config (any MCP-compatible client):

{
  "mcpServers": {
    "mdpilot": {
      "command": "npx",
      "args": ["-y", "mdpilot-mcp"],
      "env": { "GROQ_API_KEY": "gsk_..." }
    }
  }
}

Accepted keys (tried in order — Groq and NVIDIA are free): GROQ_API_KEY · NVIDIA_API_KEY · ANTHROPIC_API_KEY · OPENAI_API_KEY

---

10 MCP tools

Tool	What it does
analyze_project	Scan repo — detects stack, scripts, package manager, structure, MCP servers
generate_md_file	Generate any instruction file grounded in real repo data
generate_task_file	Turn a ticket or thread into a structured TASK.md
explain_code	Generate WALKTHROUGH.md for any file or directory
optimize_markdown	Run the 5-pass optimizer (+ opt-in aggressive 6th pass) on any markdown
image_to_prompt	Analyze an image → recreation prompt
check_drift	Detect stale docs — broken commands, broken paths, removed MCP servers
update_docs	Patch only the stale sections, preserve everything else
save_context	Persist session state to CONTEXT.md (last 5 sessions, secret-redacted)
load_context	Load prior session context + annotate stale references inline

Session memory quick-start — your data never leaves your machine:

# End of session:
"Use mdpilot to save our session context — decisions: X, next steps: Y"

# Start of next session:
"Load the project context via mdpilot"

---

Multi-model

Provider	Model	Notes
Anthropic	claude-sonnet-4-6	Best instruction-file quality
Groq	llama-3.3-70b-versatile	Free tier — recommended default
NVIDIA NIM	meta/llama-3.3-70b-instruct	Free tier
OpenAI	gpt-4o
Google	gemini-2.0-flash

---

Quick local setup

git clone https://github.com/get-mdpilot/MDPilot-mcp
cd MDPilot-mcp/mdpilot
npm install
cp .env.example .env.local   # add at least one AI key
npm run dev                   # http://localhost:3000

MCP server:

cd packages/mcp
npm install && npm run build
npm run inspect               # MCP Inspector — test tools interactively

See DEVELOPMENT.md for full architecture, file trees, environment variables, and contribution guide.

---

Stack

Next.js 16 · TypeScript · Tailwind CSS v4 · Supabase · @modelcontextprotocol/sdk · repomix · js-tiktoken · zod

---

Legal

Operated by Viveon Gizit Pvt Ltd · Bengaluru, India — Product Owner: Mohan
Privacy Policy · Terms of Service · Report an issue

MIT license