AI Founder OS — Turn Notion into a Multi-Agent Startup Command Center

This is a submission for the Notion MCP Challenge

What I Built

Every founder has the same dirty little secret: a Notion database called "Startup Ideas" where dreams go to die.

We get an idea. We write it down. But then comes the grueling part: opening 50 browser tabs to research competitors, mapping out feature matrices, digging through SaaS pricing pages, and trying to validate if the idea even makes sense. The paralysis of analysis hits hard. Instead of building, we drown in research.

I built AI Founder OS out of pure frustration. I wanted a system where I could drop a raw, messy idea into my Notion workspace, change a status property to Run, and have an entire team of AI agents do the heavy lifting for me.

Powered by Gemini 2.5 Flash and Tavily, AI Founder OS orchestrates four distinct agents to produce:

• Competitive intelligence — grounded on real web search evidence.
• Product roadmap — a prioritized feature plan with complexity ratings.
• Marketing strategy — channel-specific campaigns and content ideas.
• Viability scorecard — a 0–100 mathematical score with risk flags and an executive summary.

Everything writes back autonomously to structured Notion databases. No more "idea graveyards." It's a living, queryable startup command center.

(Plus, it's polyglot: Write your idea in Spanish, French, or Japanese. The system auto-detects the script and returns all AI-generated research in that exact language).

Video Demo

(Agrega aquí el enlace a tu video de demostración en YouTube, Loom, o sube tu video directamente a DEV.to)

Show us the code

Sherman95 / ai-founder-os

AI Founder OS - Notion Run Center Backend (v0.4.0 Agentic Run Audit)

AI Founder OS is a production-ready basic backend that uses Notion as UI + database + memory The system polls a Notion database for startup ideas with Status Run or Queued, executes a multi-agent workflow, and writes outputs into Notion databases:

• Competitor Research
• Product Roadmap
• Marketing Strategy

Competitor research is grounded on real web search evidence (when enabled) and then structured by Gemini The workflow now supports an auditable Run Center model in Notion with optional entities:

• Runs
• Evidence
• Claims
• Feature Matrix
• Competitor Scorecards

It also writes a startup viability score and executive summary back to the idea.

This release includes MCP support in two directions:

• Provider mode switch for Notion access: NOTION_MODE=api|mcp
• Founder OS MCP server exposing high-level tools (founder.*)
• Human-in-the-loop corrections loop for output rows flagged in Notion
• Run operations: Retry, Replay and…

View on GitHub

System Architecture & Flow

To make this work seamlessly without building a custom frontend, I used Notion as both my Database layer and my UI layer.

The System Flow:

1. Trigger: The system polls the Startup Ideas database for ideas entering the Run state.
2. Analysis: The Idea Analyzer Agent extracts core concepts and detects the language.
3. Research: The Market Research Agent executes live web searches (via Tavily) to find competitors and actual market evidence.
4. Planning: The Product Planner Agent generates an actionable MVP roadmap.
5. Growth: The Marketing Agent generates GTM campaigns.
6. Persistence: The notionWrites module idempotently upserts the formatted results into their respective Notion databases.

How I Used Notion MCP

When I started building this, the problem was getting the AI to interact with Notion securely and predictably. AI Founder OS integrates Notion MCP in two directions, turning a simple workspace into a full-stack operating system:

1. As an MCP Consumer

Instead of wrestling with static API keys, the backend consumes the Notion MCP hosted endpoint (NOTION_MODE=mcp) for OAuth-based access. It securely reads the "Startup Ideas" database, discovers properties dynamically, and injects the context directly into the AI pipeline.

2. As an MCP Provider

The Node.js backend exposes its own MCP server (supporting HTTP and stdio transports for VS Code/Claude Desktop clients). This exposes 10 high-level JSON-RPC tools to any MCP environment.

Key tools include:

• founder.inspect_schemas: Validates that your Notion databases have the required properties before execution.
• founder.run_idea: Executes the 4-agent multi-step pipeline.
• founder.claim_idea: Uses Notion properties as a distributed lock to prevent duplicate agent runs.

The Magic: Human-in-the-Loop Corrections

AI hallucinations happen. But because Notion is the UI, fixing them is seamless.
If an agent hallucinates a competitor's feature:

1. I review the row directly in Notion.
2. I check the Needs Review checkbox and type my feedback in the Correction Notes property (e.g., "Actually, this competitor only offers this on the Enterprise plan").
3. I trigger the founder.apply_corrections MCP tool.
4. The system re-analyzes only that specific row and safely updates Notion.

The "WOW Mode": Absolute Traceability

Any AI can generate text. But an enterprise-grade Command Center needs traceability. By enabling NOTION_UI_WOW_MODE=true, the system creates a forensic audit trail across 5 extra Notion databases:

• Runs: Execution logs, durations, and error tracking.
• Evidence: Raw web search results (URLs, domains, and AI confidence scores) proving why the AI made a claim.
• Claims: Competitor marketing claims cross-referenced with the evidence to output verdicts (supported, unknown, contradicts).
• Scorecards: Mathematical viability scoring (0-100) penalizing hidden pricing or lack of traction.

Conclusion

By leveraging Notion MCP, I didn't just build an AI wrapper. I built a resilient, multi-agent operating system that lives entirely where founders already work: inside Notion.