Nokos

Hey! New article on how Nokos uses MCP (Model Context Protocol) to give AI tools persistent memory.

The problem: Claude Code, ChatGPT, Cursor — they all start fresh every session. No memory of what you built last week.

Nokos fixes this with 3 MCP tools:
- search_nokos — search your memos + sessions by keyword or meaning
- get_nokos — fetch full content of any item
- save_session — save the current conversation

In practice: you ask Claude Code "How did I implement auth?" and it searches your actual work history — not training data, not hallucinations, but your real sessions and notes.

Two servers: local (stdio, for Claude Code/Codex) and remote (HTTPS + OAuth, for claude.ai). Same 3 tools, same protocol.

The loop: AI generates knowledge → Nokos captures it → AI retrieves it next time. Each cycle makes your AI more useful.

Full article:
https://dev.to/tomokiikeda/i-gave-claude-code-a-memory-heres-how-mcp-connects-ai-tools-to-your-knowledge-base-3l56

Do you use MCP with any of your tools? What's your experience been?

Hey! New article on the Nokos session capture architecture.

The problem: every AI coding tool — Claude Code, Cursor, Copilot, ChatGPT — stores conversations differently. And most of it vanishes when you close the terminal.

I built a pipeline that captures all of them automatically:
- Session hooks (Claude Code SessionEnd + PreCompact, Codex notify)
- VS Code extension (Roo Code, Cline)
- Cron-based file watching as a safety net for long-running sessions
- 7 dedicated parsers that normalize every format into one schema
- AI summarization so you get searchable titles and tags, not raw JSONL

The fun part: Claude Code fires "Compacting conversation" 3-5 times per session, each triggering the hook. Solved it with daily session IDs — same-day pushes upsert, next day starts fresh.

Full deep dive with architecture diagrams and code:
https://dev.to/tomokiikeda/how-i-auto-capture-coding-sessions-from-25-ai-tools-architecture-deep-dive-2p3i

Are you capturing your AI coding sessions? How do you keep track of what you've built?

Hey! Quick update on Nokos.

When you have AI writing all your code (Claude Code writes everything for Nokos), authorization bugs are inevitable. A missing WHERE user_id = ... somewhere, and one user sees another's data.

My safety net: PostgreSQL Row-Level Security (RLS). The database enforces "users can only see their own data" on every query — even if the application code forgets.

It already caught a real bug: an async embedding update ran outside our auth transaction. Without RLS, the query would have "worked" with no authorization check. With RLS, it silently matched zero rows. Bug showed up in monitoring, fixed in 5 minutes.

The key setting: FORCE ROW LEVEL SECURITY — applies policies even to the table owner (which is what most ORMs connect as). Default deny. If code skips the auth wrapper, it gets nothing instead of leaking data.

If you're building multi-tenant SaaS — especially with AI writing your queries — this is the cheapest security layer you can add.

Full breakdown:
https://dev.to/tomokiikeda/postgresql-row-level-security-saved-my-saas-from-bugs-i-didnt-know-i-had-1bb5

Anyone else using RLS? Did it catch anything unexpected?

Hey! Quick update on Nokos.

When I first built the product, I used Claude (Anthropic) for all AI features. The problem: I was losing money on every paying user. Plus plan was ¥480/month, but per-user AI costs were ~¥196.

The fix: keep Claude Code for writing the application code (where quality matters most), but switch all production AI features to Gemini Flash (where speed and cost matter more).

Result:
- Chat queries: ¥4.5 → ¥0.07 per call
- Per-user costs dropped 3-7x across all plans
- Break-even went from ~700 users to ~150 users
- Free users can now get real AI features without killing my margins

The lesson: the model that builds your product and the model that runs it don't need to be the same.

Full breakdown:
https://dev.to/tomokiikeda/claude-writes-the-code-gemini-runs-it-how-two-competing-ais-cut-my-saas-costs-by-30x-2hn3

Anyone else running multiple AI providers? How do you decide which model goes where?

Hey IH! I'm Tomoki, a solo maker from Japan.

I just launched Nokos (nokos.ai) — an AI note-taking app that auto-captures your coding sessions from 25+ AI tools.

The twist: I didn't write a single line of code. Claude Code wrote everything. Gemini powers the AI.

What it does:

• Auto-captures sessions from Claude Code, Cline, GitHub Copilot, Roo Code, and more

• Turns them into searchable knowledge

• AI generates daily diaries and reports from your memos

Built in ~30 days. Planning to launch on Product Hunt March 31.

Full story:
https://dev.to/tomokiikeda/zero-lines-of-code-how-claude-code-and-gemini-built-my-saas-35cn

Would you use something like this?