Hard decisions, weighed by a council of AIs.

Give cogitate a problem and the facts. It runs three or four LLMs in parallel — each playing a different role — then has another model attack the weakest argument from each. A final synthesizer writes a decision document that survived adversarial review.

Open sign-up. You bring your own LLM API keys — cogitate orchestrates the council.

Why a council, not just one model

One model gives you one answer, in the voice of one training set. A council surfaces what a single answer hides — second-order effects, sunk-cost traps, framings the conventional answer skipped. cogitate runs models from OpenAI, Anthropic, Google, and xAI side-by-side so a single vendor's blind spot doesn't become yours, then forces them to attack each other's reasoning before producing the final decision.

How a run works

1. You provide the facts

Paste a brief, or dictate it from your phone. cogitate auto-tidies voice-dictated text using Claude Haiku — numbers and proper nouns are preserved exactly, ambiguities are flagged, nothing new is added or removed. The raw original is kept for audit.

2. The council deliberates in parallel

Three or four personas write their position simultaneously, each grounded in your facts. Typical roles: Optimist (steel-man the upside), Pessimist (steel-man the downside), First-principles (ignore conventional framings), plus pack-specific experts.

3. A Killer attacks the weakest claim

A separate LLM tears into the softest point of each persona's answer. Specific. Blunt. Evidence-bound. This is where casual consensus dies and real disagreements surface.

4. The Scribe synthesizes

A final pass writes the decision document: the position, why, what survived adversarial review, what risks the Pessimist surfaced that you should accept, and which framings were discarded.

You can run multiple rounds. After a run finishes, you can reply with more information and trigger a follow-up — the new round inherits the prior synthesis as context.

What people use it for

Each example uses a different bundled persona pack. Pick the one that matches your problem when you start a project, or write your own pack in YAML.

Big life decisions

"Should we leave Australia and become tax resident elsewhere — and if so, where, and how?" The expat-relocation pack runs four expert personas (tax, immigration, family-impact, financial planner) plus an adversarial pass that tries to talk you out of it.

Concentrated investment positions

"I'm 80% in one stock after a 5x. Stay, diversify, restructure?" The investment-strategy pack debates the Concentrate-and-Hold case against Diversify-Now and Hedge-and-Wait, grounded in your actual numbers and tax position.

General should-we-or-shouldn't-we

"Should we hire a second engineer or stay solo for another six months?" The decision pack is the lightweight default: Optimist, Pessimist, First-principles, one rotating Killer, an Observer, a Scribe. Fast and cheap.

System-design review

"Here's the proposed architecture — what's wrong with it?" The system-design pack runs an Architect, an Implementer, Security, and Operations in parallel, with two rotating Killers looking for the failure modes none of them saw.

Asking the right questions first

"I think I want to relocate but I don't even know what to consider." The relocation-intake pack flips the script: instead of producing an answer, it produces the structured questionnaire you should fill out before the council can answer well.

Two ways in

Web: this site. Mobile-friendly — you can dictate the brief from your phone and check the result later. Optimised for long-form output that you read on a real screen.

MCP: cogitate is also a Model Context Protocol server. Drop the URL and your bearer token into Cursor, Claude Desktop, or any MCP-aware client and you can run councils, read transcripts, and manage projects from inside your editor or chat.

What you'll need

An email address. Open sign-up — we email you a one-click link.
Your own API keys for the LLMs the pack uses (OpenAI, Anthropic, Google, xAI). Stored encrypted at rest with AES-256-GCM and never decrypted for display. Cogitate orchestrates; you pay the model bill directly.
A typical run is a few cents to a couple of dollars depending on pack and rounds. The default decision pack is the cheapest; full multi-round expert councils with reasoning models cost more.

Pricing & fine print

Free to use during the beta. Once the rough edges are sanded down and there's a clearer picture of how people actually use this, cogitate may move to a paid model — most likely a small subscription that covers the orchestration, persistence, and the curated persona packs. The LLM bill stays yours either way (you bring the keys; we orchestrate). You'll get plenty of warning before any pricing changes touch your account.

Bugs are likely. Your council transcripts are private to your account, but I (the operator) can read them — if that's a problem for you, don't use it for things you wouldn't share with a single trusted human. There is a per-user daily run cap so a runaway loop can't drain your wallet, and a per-IP magic-link rate limit on the email side. Both are intentionally low for launch and will loosen as I see how it's used.

Already have an MCP API key? Configure your client with Authorization: Bearer cog_live_... against http://www.cogitate.one/mcp.

cogitate · build f714a0d3