Finish the apps you almost shipped.
Describe a feature in one sentence. Klio writes the stories, your coding agents build them by ID and show their work on every criterion, and every project reopens where you stopped.
Works with Claude Code, Codex, Cursor, Windsurf, VS Code, Claude.ai and ChatGPT through MCP. No files to export.Vibecoding is fast. Coming back to a project is slow.
You build in bursts, switch apps when a new idea hits, and lose the thread. Klio keeps the thread for every project you start.
- Login flow 80% done on
feature/login. Wrong-password state left. - New idea: an AI journaling app. Opens a fresh repo.
- Study group asks for a Discord bot. Two evenings gone.
- Back to the first app. Which branch? Which prompt? Was Resend ever set up in production?
- Twenty minutes of digging before a single line of code.
POC-33 Login form and session handling
- Next
- Wrong password shows a generic inline error
- Files
app/(auth)/login/page.tsx- Stack
- Supabase Auth · Resend key missing in preview
- Then
- POC-34 Email verification · POC-36 Route guard
One sentence in. A scoped feature out.
Klio reads the project's stack and existing features, then writes small stories your agent can finish one at a time.
- Acceptance criteria on every story, so done means done
- Likely files and the services each story touches
- You review and untick before anything lands on the board
Done comes with proof.
Agents attach evidence to each acceptance criterion they tick: test output with its exit code, a commit, a pull request. CI checks on the story's PR verify it. An agent can never mark its own work verified.
- Evidence on every criterion: test output, commits, pull requests and links
- Passing CI on the pull request verifies the test-backed criteria
- Require evidence, and a criterion can't be ticked without it
Every project opens where you left it.
A bookmark for each project: the story in progress, the criteria still open, your last note, and what comes after. One tap copies all of it as a prompt.
- All projects on one screen, with how long each has been idle
- AI rewrites a messy note into the next concrete step
It remembers what you decided.
Decisions, conventions, gotchas and commands live with the project. Every agent reads them when it starts, so nobody relearns that the staging database is off limits.
- Agents add what they learn as they work; you edit, pin and delete
- Your own preferences follow you to every project
- Copy it out as an AGENTS.md section anytime
Know what every project runs on.
Frameworks, databases, hosting, email, payments, AI APIs, domains and the tools you build with. Paste a package.json and Klio fills most of it in.
- Env var names per environment. Values never leave your machine.
- Monthly spend, and the domains or plans that renew soon
- Stories link to the services they touch
Your agent reads the board. You say the number.
Klio runs an MCP server. Connect Claude Code, Cursor, Codex, Windsurf or VS Code with one command, or sign in from Claude.ai and ChatGPT. Then build POC-F4 is the whole prompt. No Markdown files to export or keep in sync.
- Agents mark stories in progress, tick criteria with evidence and post summaries while they work
- Several agents at once: each claims its own story, and a story another agent holds is refused
- GitHub PRs that mention POC-33 move it to review, passing CI verifies it, and the merge moves it to done
From idea to shipped, without losing the thread.
Start a project
Name it, pick a starter stack or paste package.json.
Pocketbook · Next.js + SupabaseDescribe a feature
AI splits it into stories with criteria and files.
“Add login with email”Hand it off
Your agent builds it by ID and attaches proof.
› build POC-F4Come back anytime
The project opens on the story you were in.
Continue POC-33 · 80%A full issue tracker,
minus the ceremony.
The story, the agent, the proof, the next step. One shared place to keep the work moving.
Every story carries its context.
Acceptance criteria, agent claims, review notes and human handoffs stay with the work. Track environment variable names without storing secrets.
Explore the MCP toolsQuestions builders ask.
How does Klio know the work is really done?
Agents attach evidence to each acceptance criterion: test or command output with its exit code, a commit, a pull request or a link. CI checks passing on the story's pull request verify the test-backed criteria, a merged PR verifies the commit-backed ones, and you can verify anything yourself. An agent can never mark its own work verified.
Which agents work with Klio?
Any MCP client: Claude Code, Codex, Cursor, Windsurf, VS Code with Copilot, Zed, Gemini CLI, Kiro, Cline, opencode and more. Claude.ai and ChatGPT connect by signing in, with no token to copy.
Can several agents work on one project at once?
Yes. Each agent claims its own story; a story another agent holds is refused, dependencies keep the order, and you can hand a whole feature to one agent.
Does Klio see my code?
No. Klio keeps what you and your agents put on the board: stories, criteria, evidence, notes, service names and env var names. Your agent reads your code locally, the way it already does.
Where do my env var values go?
Nowhere. Klio tracks variable names and which environments have them set. A pasted value is dropped before it's saved, and evidence output is scrubbed of anything that looks like a secret.
Do agents count as seats?
No. Free covers one person with three projects and two agent connections. Pro removes the limits for one person, and Team is priced per person. Agents are never seats.
Why the name?
Clio was the Greek muse of history. Klio keeps the history of each project so your agents don't lose it.