ask-yoni

A public MCP server for semantic search over Yoni Rechtman's writing on 99% Derisible.

Posts on 99D often package multiple distinct topics in a single issue. ask-yoni splits each issue into sections, embeds each section, and lets you find the right idea instead of the whole post.

Use it

Add this URL as a custom connector in Claude (web, desktop, or Code):

https://yoni.fyi/mcp

It exposes four tools:

Tool	What it does
search_99d(query, limit=5, include_meta=false)	Semantic search at the section level. Returns the best matching sections across the archive, each with the parent issue's title/url/date.
get_post(slug)	Fetch a full Substack issue by slug, including its list of sections.
get_section(slug, section_idx)	Fetch a single section of an issue.
list_recent(limit=10)	List the most recent issues, newest first.

Try things like "what does Yoni think about AI services companies", "narrative warfare in VC", or "how should I think about seed pricing".

How it works

Substack RSS ──► Markdown ──► substack_posts ──┐
                                               ├──► sections classifier (gpt-4o-mini)
                                               │       │
                                               │       ▼
                                               │   substack_sections (kind: main/meta)
                                               │       │
                                               │       ▼
                                               │   chunker + embedder (text-embedding-3-small)
                                               │       │
                                               │       ▼
                                               └──► substack_chunks (pgvector, HNSW)
                                                       │
                                                       ▼
                                                 MCP server (FastMCP, Streamable HTTP)
                                                       │
                                                       ▼
                                                 https://yoni.fyi/mcp

• Storage: DigitalOcean Managed Postgres with pgvector. Tables: substack_posts → substack_sections → substack_chunks. Sections classified as meta (link lists, sign-offs, unrelated promos) are filtered from search by default.
• Embeddings: OpenAI text-embedding-3-small (1536-dim) over ~500-token paragraph-aware chunks, each prefixed with # {post_title} — {section_title} for retrieval context.
• Section classification: A single gpt-4o-mini call per post classifies each section (main vs meta) and writes a one-line summary.
• Hosting: A small FastMCP/FastAPI server runs on the existing DigitalOcean droplet behind nginx + Let's Encrypt. Deploy is git push origin main — the sync cron picks code up on its next tick. Read CLAUDE.md for operational details.
• Sync: A cron job hits the Substack RSS feed twice daily (17:00 + 03:00 UTC — primary + backup, spanning the typical Friday-morning publish window). New posts are split, classified, embedded, and inserted. Each run's status, new-post count, and any failed slugs land in substack_sync_runs.

Local development

# 1) Create venv (Python 3.12+) and install deps
uv venv --python 3.12 .venv
.venv/bin/pip install -r requirements.txt
.venv/bin/pip install mcp fastapi 'uvicorn[standard]'

# 2) .env at the repo root
cat <<EOF > .env
DATABASE_URL=postgres://...
OPENAI_API_KEY=sk-...
EOF

# 3) Run the server
.venv/bin/uvicorn app.server:app --host 127.0.0.1 --port 8091

# 4) Exercise it
.venv/bin/python scripts/test_mcp.py

Source

• Newsletter: https://99d.substack.com
• Repo: https://github.com/yrechtman/ask-yoni