(fractal + task)
Get shit done with your AI partner.
Five rules. One open-source tool.
Fractask is a way to get things done with your AI — built for the hybrid world where humans and bots work side by side, not one behind the other.
Start with a goal. Anything you actually want to do — plan a wedding, ship a website, hire your first employee, move countries.
You and your AI break the goal into smaller parts. Then break those into smaller parts. Keep going until each part is something one person can finish in one sitting. Three at a time keeps it readable.
Anyone in the duo picks a part and does it. Human or bot. Notes, files, and links live on the part itself, so the next session — human or agent — picks up cold.
Together, you build something real.
A fractal is a pattern that repeats at every level. Zoom in on a coastline and you see smaller coastlines. Zoom in on a tree and you see smaller branching trees. Zoom in on a snowflake and you see smaller snowflakes.
The same rule, applied again and again, builds infinite complexity out of one simple instruction.
Work breaks down the same way. A goal contains projects. A project contains tasks. A task contains sub-tasks. The same rule — break by ~3, until every part is actionable — applies at every depth.
That’s why it’s called Fractask.
The Polish-French-American mathematician who coined the word fractal in 1975. He saw self-similarity everywhere — in cotton prices, mountain ridges, coastlines, lightning bolts, the human lung. Where most mathematicians saw mess, he saw patterns nested inside patterns.
His most famous discovery — the Mandelbrot Set — is the canonical fractal. Zoom in forever and you keep finding the same intricate detail. One equation, infinite depth.
“Clouds are not spheres, mountains are not cones, coastlines are not circles, and bark is not smooth, nor does lightning travel in a straight line.”
He gave us the language to see the world as it actually is: recursive, repeating, nested. Fractask carries that vision into the work you do every day.


One sentence. Anything. Fractask doesn’t care if it’s big or small.
Your AI proposes parts; you accept, edit, or rewrite. Three branches, three sub-branches, as deep as you need.
The bot picks the boring parts. You pick the hard ones. Or vice-versa. Either of you can claim any part.
A doc, a Loom, a Slack thread, a screenshot — all live on the part. Anyone opening it later starts informed.
Mark it done. Move on. The tree shrinks. The thing exists.
One root. The thing you actually want. Everything grows downward from here.
Decomposition is a conversation, not a monologue. You and your AI split it together — until every part is something you can actually do.
Aim for ~3 sub-parts at each level. Three is what a brain can hold. Go as deep as the work needs.
Human or bot. Pull the part you can do; let the other one pull what they can. Pulled work has an owner.
Notes, files, links — they belong on the task itself. The next session picks up cold.
Review is a tool, not a rule. Use it when a part is risky or you want a second pair of eyes. Most parts ship without it.
MIT-licensed. Runs on your laptop or any server. Works with Claude Code, Cursor, Claude Desktop, and any MCP-compatible agent. CLI, web, and MCP all share one task tree.
~/projects/launch › pnpm install && pnpm --filter fractask/cli build ✓ ready ~/projects/launch › fractask add "ship the new homepage" + created 9rEH4P ship the new homepage ~/projects/launch › fractask ls --tree ○ 9rEH4P ship the new homepage ○ 7Lm0Wq hero copy ○ 8oNxRb 3D animation ○ 6PqdSe open-source feel
Or run it as an MCP server. Your agent calls create_task; the result lands in the same tree your CLI sees.

The Fractask method in 30 pages. Where the idea came from (Kasparov, 1998 · Miller, 1956). What GTD got right. What it missed in the age of AI. A worked walkthrough. A short call to action.
Three places agents already live. Each one becomes a node on the same tree your CLI, MCP, and web app see.
Forward a message, mention the bot in a group — every chat thread becomes a tree on the side. Replies update the task. The thread stays the trail.
Read the page→Hand a goal to a Hermes agent. It picks up tomorrow exactly where it left off, on the same tree — not a fresh tool loop with no memory.
Read the page→OpenClaw drives the browser. Fractask holds the plan. Connect the two MCPs and computer-use finally has a memory.
Read the page→Everything in the OSS, plus the things only a team needs. We onboard each team by hand.
Share goals across the org. Owners, members, and bots all see the same tree.
Ping a part to chat and it lands as a block. Reply to update. The thread is the trail.
Run an agent every morning. It picks up the parts it can do and reports back.
Google login. Daily backups. We host it; you stay out of the way of the work.
Early access. We onboard each team by hand.