Describe what you want to build. Specky asks questions, detects contradictions, and records every decision — who, when, why. The result: a versioned system plan for your dev team, or the input for Claude Code to build it step by step.
7 days free · no card · your specs never train AI models.
The architecture in one chat, the data model in another, a scope change in the fifth. Two weeks later, nobody knows where the current version lives.
"Why did we pick PostgreSQL over Mongo?" The answer is somewhere in the scroll. A chat doesn't remember who decided, when, or what was rejected.
You paste fragments from memory into your agent. Code gets built from what you remember — not from what was actually decided.
Chat is great for thinking. Terrible as the source of truth about your project.
| ChatGPT / Claude / Gemini | Specky | |
|---|---|---|
| Decision history | scattered across chats | one project, one source of truth |
| Conversation context | you paste fragments | the model sees your whole project specification |
| Versioning | none | specification snapshots with history for every requirement |
| Decisions | lost in the scroll | who, when, why — saved on the requirement |
| Consistency | the model doesn't see the whole | contradiction detection across requirements |
| Teamwork | copy-paste links | shared project, roles, joint resolution |
| Agent connection | none — copy-paste | MCP server: the agent reads the spec, proposes changes, reports progress |
| Implementation state | "did we do that?" | status of every requirement: open → code ready → done / to fix |
Specky is the layer chats are missing — it completes your stack, it doesn't replace it.
You resolve a question — Specky records the decision on the requirement: who, when, what the alternative was. The requirement's history shows how it reached its current shape. Come back a month later and you know why, not just what.
Requirement 3.2 says "export to PDF". Requirement 5.1 says "CSV only". A chat won't catch it — it doesn't hold the project as a whole. Specky points to both sources and asks you to resolve it.
AI proposes, the human decides. No change without your approval.
One command and Claude Code reads requirements straight from Specky — with acceptance criteria and full context. The agent can also propose changes: they go to your review queue, never straight into the spec.
When the agent finishes an implementation, the requirement gets status code ready. You test and mark it done — or to fix with a note the agent picks up as a task. Specky always knows what already exists in the code.
The agent reports it's finished. You're the one who says done.
Drop in a call transcript or your raw meeting notes. Specky reads the whole thing and works out where it belongs in the project — new requirements, agenda points to decide, acceptance criteria — each proposal pointing back to what was actually said.
Nothing enters the spec on its own. You approve each proposal, edit it, or dismiss it.
Specky is the memory and discipline your agent lacks: one coherent spec instead of forty chats, markdown export, MCP in your terminal.
Requirements, client decisions, and choices in one place — with roles and attribution. No more "who decided this, and where".
Start a conversation about your project — like ChatGPT or Claude, but focused on your spec. Or upload docs (PDF, DOCX, MD). Requirements emerge as proposals: you accept or reject each one.
Specky asks questions and flags contradictions between requirements. Talk each one through — with your full spec as context, not a pasted fragment. The moment you decide, it's saved as a decision on the requirement.
A coherent set of requirements with acceptance criteria, captured as a snapshot. Whoever asks — team, client, or agent — you always know which version is current.
Export the plan, or feed Claude Code (spec + starter prompt). The agent ships requirement by requirement and reports code ready; you mark done or to fix. New feature → diff → next version.
The specification doesn't end at v1. Specky carries it through the whole life of the project — and knows what of it already exists in the code.
Data hosted in the EU (Hetzner). Your files stay on EU servers under GDPR.
Your specs never train AI models. Paid APIs only, with training excluded.
Every AI proposal is justified — with a cited source you can check.
7 days free. No card.