Everybody is talking about vibe coding. People have been vibe product managing for years.

Meridian is where the argument about what to build next gets tested instead of won - a claim written down before anyone looks, a verdict at the end.

Start an investigationSee how it works
The Meridian canvas mid-investigation: a hypothesis on the left, the Onboarding funnel workstream holding its three tasks, and the evidence and insight linked below it.

Nobody ever writes the question down

Every roadmap fight has a question under it, and nobody wrote it down. So every new fact feels like it changes everything. The notes are scattered - Apple Notes, a wiki nobody reads, the email thread where you are still digging for what Joe said on last week's customer call. With nothing written down, the roadmap goes to whoever needs it most - a salesperson closing one low-margin deal. Meridian is where the question goes.

Open a question. Close it with a verdict.

A near-empty canvas holding a single hypothesis card, with the add-item menu open showing what else can go on it: evidence, insight, conclusion.

1 · Pose a hypothesis

Write down what you think is true before you look. It sits on the canvas where the whole team can watch it hold up or not.

A close-up of the evidence behind one workstream: three cards naming their sources, linked to the insight they support.

2 · Gather evidence

Bring what you actually have - the interview notes, the funnel number. Link each piece to the claim it speaks to, including the evidence that says you are wrong.

A concluded investigation ending in a verdict card reading that the benefit-led headline lifted trial starts, with the insights it produced beside it.

3 · Reach a verdict

The question closes in public with the reasoning still wired underneath. Ask again later and the answer has a URL.

Prima

An agent that already read the project

Ask for a status update and Prima reads the evidence on the canvas before she answers. Her work goes back onto the same canvas with her name on it - you can watch her cursor move while she works, and check later exactly what she did.

Prima's chat panel docked over the canvas, showing a tool step reading the project's canvas context and a drafted status update citing the evidence on screen.

1 · Bring the agent you already pay for

Meridian speaks Model Context Protocol, so Claude, Codex and Cursor connect straight into your projects with a self-service API key. No waitlist.

2 · Prima can never do more than you can

She acts on your behalf, inside your permissions, and the database stops her at the edge of them. She cannot open a project you weren't invited to. Some things stay human-only - deleting a project is one of them.

3 · Another organization can never see your work

Enforced in the database on every request. A page cannot show what the database never returns.

What Meridian will never do

  • Break work into tickets. Linear does engineering tickets, code review, and now pointing agents at them. Meridian will never be the place where Terry picks up ASS-123 and Sally picks up ASS-124. The verdict you reach here might be the reason those tickets get written. Writing them is Linear's job.
  • Build a coding agent. Linear is doing it, Vercel is doing it. Claude Code and Codex do that work better than anything we could bolt into a sidebar, so Meridian sends your verdicts out through the same Model Context Protocol connection, to the agent you already pay for.
  • Hold the big strategy docs. Those are real work, and they live in Notion. Meridian links out to them.
  • Sell templated chart clipart. A picture of a funnel is not evidence, and Miro will happily sell you a whole drawer of diagram furniture.

FAQ

The first idea did not survive its own process

Before this I was head of product at a Series C startup selling to Fortune 500 banks, watching roadmap decisions get made on what the CEO wanted, with nothing from a customer underneath them.

Meridian started in July 2024 as Component Monkey, a Figma companion for turning design components into React code. I watched people less technical than me do that exact work by talking straight to an LLM, and the evidence said the layer I was building was not needed. So I killed it. Roughly six months and a few thousand dollars of my own money in, what survived was the process: write down what you believe, put the evidence next to it, and change direction when the two disagree.

Pick the argument that never stays settled

Open it as an investigation and close it with a verdict. Whoever reopens it later has to argue with the evidence first.

Start an investigation