Specbench

Quick start

From sign-in to your first spec change

This walkthrough uses the hosted app at app.specbench.io. Self-hosting instead? The product is identical once it's running — do the install first, then come back here.

1. Create your organization

Sign in and create an organization. The organization owns everything — projects, teams, and the people you invite. Invite teammates from the Users page whenever you're ready; nothing below needs them yet.

2. Create a project

A project is one product's spec. Create one from the Projects page and give it the product's name, not the team's — projects outlive team structures.

3. Open a workstream

Every spec edit happens inside a workstream — a long-lived thread for an area of work, like "Checkout flow" or "Identity & access". Create your first one and give it a framing: what it's for, what's in scope, what's explicitly not.

You don't need many. A workstream is a place work returns to, not a ticket.

4. Open a Task

Inside the workstream, open a Task for the first concrete slice — the thing you'd want agreed and shipped as one unit. Tasks are where changes accumulate, get reviewed, and merge onto the project's shared spec when marked done.

5. Author your first artefacts

Start with the vocabulary and shape of the domain:

  • Bounded Contexts — the major areas of your product, each with its own spec prose.
  • Roles — who uses the product, what they need, where it hurts today.
  • Glossary Terms — the words your team uses, defined once, with the near-misses to avoid.

There's no required order. Most teams start by naming two or three contexts and the terms they keep arguing about.

6. Connect an agent

Specbench exposes its whole authoring surface over MCP, so an AI agent can read and edit the spec the same way you do — attributed, inside a workstream, landing on a Task. Point your agent at the MCP server and sign in when it asks; Agents & MCP has the per-client setup for Claude Code, Claude Desktop, Codex, and ChatGPT, plus the guided skills that make the agent a modelling partner rather than a bare toolset.

Where next

On this page