nateherkai/roast-room

This is a simulated advisory board for founders weighing a concrete business call: you give Claude Code a decision, its numbers and your options, and persona agents score the options and argue before producing a verdict and a PDF. Worth a look if you already pay for Claude Code and want a structured second opinion rather than a chat reply.

The problem

A founder wants to decide whether to raise prices, hire, or change positioning. Today that means either trusting one model's single answer or collecting scattered opinions from a few people. With this project the workflow changes early: the Chair reads the question, proposes which advisors should sit and why, and asks for missing facts like revenue, options and success criteria before anyone argues. The result is a scored verdict across the options plus a session folder with a Markdown report and downloadable PDF.

What's different

The mechanism is a set of Claude Code sub-agents, one per advisor, each backed by a locally stored "brain" compiled from that person's public essays, letters, talks and videos. The Chair runs rounds, keeps score weighted by each advisor's pillar, caps the verdict when a seat flags ruin risk, and forces disagreement when everyone agrees too fast. Because advisors are plain skill and agent files, you can add your own with /newroaster, and the docs say every verbatim quote is checked against the source brains.

The local web room only replays files the Chair writes; it makes no model calls itself, which keeps the visuals cheap and the reasoning inside Claude Code. The trade-off is cost: a full council runs three sub-agent passes per seat, 21 for seven advisors, which the README says takes a real bite out of a Claude plan. The project also frames these people as simulations, so treat the verdict as a structured prompt, not authentic advice.

Who should try it

It fits a founder or operator who already works in Claude Code, has a decision with two or three options and real numbers, and wants several viewpoints before committing. Smaller, reversible calls can use --quick, which seats four advisors for two rounds and uses far less of the plan. It is a poor fit if you expect the tool to supply market data: the council reasons only from what you give it. If you do not use Claude Code, an optional API-runner mode lets the room run councils with an Anthropic key, but that bills your API account.

Requirements and limitations

You need Claude Code and Node.js 20 or newer; the room has no other dependencies, and the council is documented to run the same in Claude Code without the room. The practical constraint is usage: each full council consumes multiple sub-agent runs, so frequent use on a paid plan needs planning. Default mode routes through your Claude plan with no API key, while the API-runner mode needs an Anthropic key and about 25 calls per full council, as described in the setup notes.

Ongoing care matters: added advisors depend on fetching public material, and the quote checker is a script you run per session rather than a guarantee about the underlying sources. Three finished example sessions ship with the repo, so you can inspect output style before spending plan usage.

HotRepo's take

HotRepo's tracked coverage begins 2026-10-24, and its only dated signal is a #11 placement on the NEW-REPOSITORY board for JavaScript that day, with 1,098 total stars. There is no comparable daily history, so persistence cannot yet be judged from these samples; the project is 14 days old. Before trying it, check whether the default seats' compiled frameworks match the decisions you actually face.