In spec-driven development, your coding agent builds to a spec you reviewed. Here you describe what you are building in a paragraph; Entalpa drafts the spec, asks what it cannot guess, and serves it to your agent over MCP in every session.
No account to try · A draft from one paragraph · One command or one click to connect your agent
Try it on your next feature
Free, no account needed. A few runs per visitor per day.
Already writing specs?
Spec Kit and OpenSpec write one spec per feature or per change. Paste a public repository and read it as one project, before you sign up. Nothing is written to your repository.
Where it stalls
The idea is sound: the agent builds to something you decided. What breaks is the spec as a markdown file, in four ways.
With spec files
It comes up in every thread where people say why they dropped spec-driven development: hours of writing and refining before any code. And where the finished spec is still silent, the agent guesses.
You write a paragraph. Entalpa drafts the stakeholders, stories and requirements, and asks only what your paragraph leaves open, before the code exists. Each answer goes into the spec.
With spec files
Each feature gets its own spec, and nothing merges them. In a hand-checked sample from 77 public Spec Kit repositories, about one later feature in three changed something an earlier spec still describes, and nothing marked the earlier one out of date.
One current spec of the whole system: stakeholders, stories and requirements, linked to each other. On import, a later feature revises the story it changed instead of adding a second copy, and where a requirement may replace an earlier one, you get a question, not a silent edit.
With spec files
The file the agent reads is a file it can write. Asked to "update the spec to match", it will, and it does it unasked too, so a rule you decided and a rule the model filled in end up looking the same.
Lock a requirement and it is never handed to the model to rewrite. Every change, yours or your agent's, lands in the project's history, where you can compare versions and undo.
With spec files
A spec describes one change, so once it ships it is a record of the past. Next session the agent starts again from a prose file and the code, and re-derives the rest.
The spec is what the system is, not what one change was. Your agent reads the part it needs over MCP in every session, and reading or writing it never spends credits.
Spec-driven development, and what happened to it: the longer version, in the words of people who tried it.
How it works with your agent
Or import the specs you already have. Entalpa drafts the stakeholders, stories and requirements, each with acceptance criteria.
Entalpa lists only what your description leaves open. Your answers go into the spec, and the model never rewrites the requirements you lock.
One command for Claude Code or Codex, one click for Cursor or Kiro, and any other MCP client works too. Nothing to paste into the chat.
It reads the requirements it needs and marks what it built as it goes, so the spec says what the system does now, with no specs folder to tidy by hand.
What we measured
≈1 in 3
later Spec Kit features changed an earlier one
1,007 feature specs from 77 public repositories, with a hand-checked sample of later features. 71% of the specs still say Status: Draft, more than half of those with every task ticked.
−48%
cost to a correct build
−57%
fix-it iterations
Controlled evaluations: the same feature and the same coding agent, once working from a short brief and once from structured Entalpa requirements, graded by the same test suite. Costs are from real agent runs, not estimates.
Spec Kit, Kiro and OpenSpec
Two of them are free, each is better than us at something, and the comparison pages say what. This is only the shape: whether there is one current spec, whether each requirement has a status, and whether the model can rewrite it.
| Tool | One current spec | Status | Traceability | Safe from AI | Round trip |
|---|---|---|---|---|---|
| GitHub Spec Kit | no, per feature | per feature | tasks → stories | no | yes |
| Kiro | no, per feature | task ticks | tasks → requirements | no | — |
| OpenSpec | specs/, archived | no | headings only | no | yes |
| Entalpa | yes, over MCP | draft → verified | stakeholder → story → requirement | yes | — |
Each tool's default, documented behaviour as of October 2026. "Safe from AI" means a requirement you lock is never handed to the model to rewrite; "Round trip" means Entalpa imports the tool's files and exports back to them.
To try it, none: describe what you are building, or paste a public Spec Kit or OpenSpec repository, and the draft comes back on the page without an account. To use it, one command (Claude Code, Codex) or one click (Cursor, Kiro) connects your agent, and from then on it reads the spec by itself in every session.
Short answer: No.
They decide how your agent works on the next change. Entalpa holds what the product must do, across all the changes. Spec Kit and OpenSpec repositories import directly, and the current spec goes back out as a Spec Kit feature or an OpenSpec folder whenever you want.
Kiro reads and writes Entalpa over MCP, like any other client.
A plan is the right tool for one change, and it is gone once the change ships. Entalpa keeps what the plan was built against: every requirement the system has to meet, which ones are decided, and which questions are still open. The agent reads it next session, and so does the next agent.
Every change, yours or your agent's, goes into the project's history. Undo steps back through the last fifty changes, and you can compare the current spec with any named version and restore it.
Whenever you want: as a Spec Kit feature, as an OpenSpec folder, as ReqIF for the requirements tools that read it, or as Markdown, CSV and JSON. Your agent can also write it into the repository over MCP.
It is free to start, with no credit card. The AI work inside Entalpa, such as drafting a project, revising it or folding an import, uses credits. Your agent reading and writing the spec over MCP never does.
Any client that speaks MCP. There are setup guides for Claude Code, Cursor, Codex, Windsurf, VS Code with Copilot, Gemini CLI, Kiro and nine more.
Start from a paragraph or from the specs you already have. Your agent reads it over MCP from the first session.
Free to start · No credit card
