The PRD nobody updated
Written once, linked everywhere, true for about a week.
The product OS
One living spec.
Your whole product runs on it.
CrystalSpec is the product OS for teams that build with AI. Strategy, specs, delivery, sales, and onboarding share one approved source of truth, and your coding agents read the same one.
The problem
Your product lives in six places.
Each one remembers a slightly different product. Somebody spends Friday reconciling them, and the coding agent never got the memo.
The strategy in a slide deck
Last quarter's market, frozen in a deck the team has stopped opening.
The tickets in a tracker
Thousands of tasks, and no way to tell which product behavior each one serves.
The deals in a CRM
Sales promising features from memory, because the spec lives somewhere else.
The answers in chat
Every “wait, how does this work?” answered again, by whoever happens to be online.
The agent guessing
A coding agent filling the gaps with confident fiction from a prompt.
How it fits together
Built like an operating system.
A versioned, human-approved kernel. Apps that read it. Drivers that carry it into the tools you already use. System calls for your agents.
System calls & drivers
How agents and other tools reach it
- Hosted MCP
- GraphQL API
- Signed webhooks
- GitHub · Linear · ClickUp tasks
- GitHub & GitLab checks
- Slack Q&A
Apps
Tools that read the published revision
Shell
Where people work with it
- Inbox
- Discussions & @Spec
- Project chat & live voice
- Cmd+K
- Prototypes
- Public share links
Kernel
The published revision
- Flows
- Data models
- Roles
- Test cases
- Phases & epics
- Glossary
- Versioned · human-approved
Define
The kernel: a spec that stays true.
Typed flows, data models, roles, and test cases, cross-referenced into one vocabulary. Edits happen in a draft revision; publishing cuts a version with field-level diffs you can revert.
- Branching flows with steps, sub-flows, and decision points
- AI proposals validated before they can apply
- Inconsistency analyzer for contradictions and gaps
Decide
Strategy that has read the spec.
Cofounder convenes a four-voice board on any question, grounded in your spec, Live SWOT, and competitor radar. Good advice becomes a decision with a test plan and a revisit date.
Meet Cofounder
Build
Delivery linked to behavior.
Native tasks in saved List and Board views, each linked to the flows it delivers. Click through generated prototypes before anyone codes, then let agents read the published revision over MCP.
- Custom statuses, labels, assignees, priorities, and discussions
- Push atomic tasks to GitHub, Linear, or ClickUp
- Paid teams check GitHub and GitLab changes against the spec
Sell
Sell what the product actually does.
A lightweight Sales CRM keeps leads and clients next to the spec. Sales Training lets reps rehearse against the real product by chat or live voice, and Pitch drafts the investor story from the active revision.
Explore Sales
Run the team
One Inbox for everything that moved.
Notifications, discussions, and assigned work arrive in one Inbox with reading panes. Threads carry replies, reactions, and @Spec, the assistant that answers from the spec. Newcomers get onboarding books, role guidelines, FAQs, and a quiz.
See onboarding
The AI layer
AI everywhere. In charge nowhere.
Every surface has an assistant that knows the project. None of them can change the spec behind your back.
Proposals, never silent edits
The assistant returns structured create, update, and delete proposals. Each is validated before it can apply, and nothing writes until a person approves it.
Talk it through
Project chat knows the live spec, and live chat lets you think out loud by voice. Changes still arrive as drafts that wait for approval.
Big changes, one review
Hand a large change request to a background change set and review everything it prepared in a single pass.
Cmd+K over everything
Jump to any flow, step, data model, term, or tool and land with the editor open.
Apps
A dozen tools, one source of context.
Every tool reads the same published revision, so the onboarding book, the sales roleplay, and the compliance report all describe the same product.
FAQ
Questions before you boot it.
What does “product OS” mean here?
One system where everything a product team does reads from the same approved source of truth. In CrystalSpec that source is the published revision of your living spec: flows, data models, roles, and test cases. Strategy, delivery tasks, sales, onboarding, and coding agents all work from it instead of keeping their own copy of the product.
Do I have to replace Jira, Linear, or my CRM?
No. The Tracker and Sales tools cover spec-linked tasks and a lightweight deal pipeline, which is plenty for many teams. Teams with an established tracker keep it: publishing pushes AI-decomposed atomic tasks to GitHub, Linear, or ClickUp, each linked back to the revision it came from.
How do coding agents fit in?
Agents connect to a hosted MCP server, verified with Claude Code, Claude Desktop, and Cursor, using scoped personal tokens. They read the published revision, list what changed, and ask the project questions. A scoped GraphQL API and HMAC-signed webhooks cover the rest.
Is any of this generated without review?
Specification changes are not. The AI proposes and a person approves every create, update, and delete, and each decision is recorded. Tool outputs such as onboarding books, guidelines, and board advice are generated from the approved spec and never edit it.
What does it cost?
Start with a permanent Free workspace with up to 5 active projects, 1 member, and 500 AI credits each month. Team is $49 per month with the first seat included, then $10 per additional seat, with one shared monthly AI pool. Spec exports, repository automation, and Smart AI need a paid plan.
Every change versioned. Every revision revertible.
Boot your product OS.
No reinstall required.
Start freeOngoing Free plan, no credit card.Human-approved AIAgents read the same spec