996.fm — On-the-record founder advice for LLMs and AI assistants

Every position is a real sentence somebody actually said

Not a summary of the internet. Here is one answer, followed all the way back to the tape — you can click through and hear it.

1Someone asks
Should I let my team work fully remote, or require people in the office?
2996.fm returns positions — not an average
4 debates · 32 positions · 27 readings · 23 named operators
Tobi Lütke · Shopify · Apr 2023 — a new company should weigh remote’s productivity discount, roughly 10–20%, against the talent density you gain by hiring anywhere.
Max Levchin · Affirm · Feb 2025 — sufficient in-person time is non-optional; anyone claiming a team never needs to meet is kidding themselves.
Both are on the record. The graph never tells you who wins — it hands your assistant the disagreement and lets it argue.
3And here is the tape it came from
20VC · Tobi Lütke · 26:31
“…even if you don’t think you can work as well together remotely, let’s put a 10, maybe 20% discount on on productivity, is the talent density more than 20% higher if you can hire everywhere than if you’re restricted onto one city.”
Verbatim, including the stumble. Every position in the graph carries a link like this one.

~6 seconds

from question to answer

under 1¢

what that answer cost to produce

32 positions

from 23 named operators, each with a date and a timestamp

3 out of 5

how that answer scored on our own eval

We grade ourselves, and we publish the grade

The answer above was scored by a separate model reading it cold, against a rubric that punishes padding and rewards admitting a gap. It got a 3 out of 5:

Dated, attributed startup tradeoffs arm the decision, but customer-meeting padding and grouping periodic togetherness with mandatory office work make the unqualified strong-coverage label overstate relevance.

In plain terms: it padded the answer with a debate about in-person sales meetings, lumped “meet up now and then” in with “be in the office”, and still called its coverage strong.

This is our own eval, run on ourselves — so here is every other question we tested, not just the one we chose to show you.

All 18 questions we tested3.1 average

1 worst

1

2

3

8← the one above

4

6

5 best

0

Nothing scored a 5 — we would not believe it if it had. The low scores are the failure we are still fighting: padding, and claiming stronger coverage than the graph actually holds.

What is 996.fm?

LLMs rely on generic, unattributed startup advice averaged from everything ever written about it.

996.fm pulls attributed, dated positions from operators on the record — as they said them on podcasts, with word-accurate, timestamped transcripts as receipts. Connect it to Claude or any LLM and get answers that name who said what, on which episode, and where they disagree.

Without 996.fm

✕Advice averaged from training data, no one accountable for it
✕“Experts say” with no names, dates or sources
✕One smooth consensus where real operators disagree
✕Quotes and anecdotes you can’t verify
✕Hours of podcasts to find the one segment that matters

With 996.fm

✓Positions attributed to a named operator, dated to the day the episode was published
✓Every position linked to the minute in a word-accurate transcript
✓Disagreement preserved — grouped by how each person reads the question
✓Free, no signup, no API key
✓One URL to connect to Claude, Claude Code, Cursor or any MCP client

Ask the question you were going to ask anyway

Once connected, your assistant calls 996.fm when a founder question comes up. Add use 996fm to force it.

1Write your question naturally
2Add use 996fm — or turn on founder-mode and skip this
3Get attributed positions, with receipts, instead of an average

No averaged advice, no unnamed “experts”, no quotes you can’t check.

We’re at $600k ARR and I still close every deal myself. When should I hand off selling to a sales team? use 996fm

996fm · find_perspectives → 4 debates, 36 positions, 26 readings

Hire before you feel ready — Carles Reina (ElevenLabs, Feb 2026): bring in a scrappy, iterative sales leader at zero revenue. 46:56 · Matt Plank (Rippling, Dec 2024): hire a GTM leader well before the traction people say you need. 54:48

Founder-led until it repeats — Sam Blond (Brex, Apr 2024): not before a handful of non-friends-and-family customers and a process you believe repeats. 14:17 · Carlos Delatorre (Harness, Jan 2025): no sales leader in the first two or three years. 6:29

…and 24 more readings. Harry Stebbings (Apr 2023): the $1M ARR trigger is meaningless. 9:10

Step 1 — connect the graph

Add 996.fm in one command

One remote MCP URL wires 996.fm into Claude, Claude Code, Cursor and every other MCP client you use. No key, no signup.

https://www.996.fm/mcp

Or let your agent do it — paste this into Claude Code, Codex or Cursor. It connects the graph, installs founder-mode, and tells you what to ask first.

Install 996.fm for me — follow https://www.996.fm/install.md

Or wire every agent on your machine at once — Claude Code, Cursor, Codex, VS Code and more:

npx -y add-mcp https://www.996.fm/mcp -g -y -n 996fm

Pick your agent

Claude.ai

Claude on web and desktop

Customize → Connectors → Add custom connector. Paste the URL and leave Authentication on “None”.

https://www.996.fm/mcp

founder-mode on Claude.ai

Optional — the connector alone now carries the consult contract, because claude.ai passes tool descriptions but drops the server’s prompt, resource and instructions. On a phone, stop after the connector. For the fuller version on desktop: Settings → Capabilities → “Code execution and file creation”, then Customize → Skills → Add → Upload skill → founder-mode.zip, or paste founder-mode.md into your project instructions.

Using a different agent? Any client that supports remote MCP over HTTP works with the URL above.

Step 2 — turn on founder-mode

The skill ships with the server

Step 1 gives your assistant the graph. founder-mode tells it how to use it: consult before advising, present the readings people actually hold instead of one smoothed answer, name and date every position, leave hedged claims hedged, and say plainly when the graph is thin rather than dressing up adjacency as evidence.

How much of it arrives on its own depends on your client — the picker above spells out which, per client. Here is the shape of it.

Comes with the connection

Claude Code, VS Code Copilot, Codex and ChatGPT read the server’s MCP instructions, so the short contract loads the moment you connect. Claude Code and VS Code go further and expose the full skill as the founder_mode prompt and the skill://founder-mode/SKILL.md resource.

Required on claude.ai

Custom connectors on claude.ai pass the tools and nothing else — no prompt, no resource, no instructions. Turn on Settings → Capabilities → “Code execution and file creation”, then upload the skill at Customize → Skills → Add → Upload skill. Without it you get the graph but not the contract.

Always-on, anywhere

No MCP client installs a skill for you. Drop the file in your client’s skills folder — ~/.claude/skills/, ~/.cursor/skills/, ~/.agents/skills/ — or pastethe contract into its rules, and it applies to every conversation without being asked.

Download founder-mode.zip Read the contract