Fractask
Integration · OpenClaw

OpenClaw drives the browser. Fractask holds the plan.

Connect both MCPs to the same agent and the loop changes shape. Each step OpenClaw takes — a click, a form, an extraction — lands on a task in the tree it can read back tomorrow morning. No more runaway browser sessions with nothing to show for them.

OpenClaw.ai →Open Fractask →
Both servers speak MCP•Works with Claude Code, Claude Desktop, Cursor•No glue code
openclaw · session 84A2
◀ ▶ ↻portal.acme.com/invoices/new
Acme Procurement — submit invoice
VendorSunbek Ltd
Invoice no.INV-7741
Amount$ 4,200.00
Submit← claw clicked
Step log
› navigated to portal.acme.com/login› typed credentials · auth ok› clicked "Submit invoice"› extracted reference INV-7741› fractask.create_task("Upload signed PO", parent=acme-onboarding)
fractask · acme onboarding
Procurement › Acme onboarding

Onboard Acme as vendor

  • Submit W-9
  • Upload signed POopenclaw · doing
  • Get tax ID confirmation
Notes

Agent log: openclaw filed invoice INV-7741 at 14:02. Awaiting countersigned PO from finance@acme.com — the claim expires 17:00.

One agent, two tools. OpenClaw fills the form on the left; Fractask records the move on the right. The next session — yours or another bot’s — picks up exactly where this one left off.
What you get

A browser agent that writes things down.

  1. 01

    Every browser action lands as a task.

    When OpenClaw clicks something irreversible — files an invoice, books a meeting, emails a vendor — it calls fractask.create_task with the artefact attached. Tomorrow morning your tree shows what got done, by whom, with the reference number.

  2. 02

    Long-running browser jobs survive disconnects.

    OpenClaw claims a Fractask part before it starts work. If the session crashes mid-flow, the part is still claimed; the next agent (or you) can resume from the last logged step instead of starting the form over.

  3. 03

    You can hand the same tree to a human.

    OpenClaw picks the boring repetitive parts — login flows, status checks, form-fills. You pick the judgement calls. Both sides write to the same tree, so the hand-off is just unclaiming a task.

  4. 04

    Every claim has a deadline.

    Fractask refuses to let an agent hold a part forever. If OpenClaw goes silent before progress lands, the part returns to the pool and a human gets pinged. No more zombie sessions.

  5. 05

    Auditability you can hand to a finance person.

    Every step OpenClaw takes is logged on the part itself — URL, screenshot reference, extracted text. When something goes sideways, the trail is on the task, not buried in a chat transcript.

How it runs

Two MCP servers,
one agent.

01

Register OpenClaw as an MCP server.

Follow the OpenClaw docs — it ships a stdio server you can attach to Claude Code, Claude Desktop, or Cursor. The agent gets browser tools: open, click, type, extract.

02

Register Fractask alongside it.

Same client, second server. Six tools — list_tasks, get_task, create_task, update_task, move_task, delete_task — pointing at the same tree your CLI sees.

03

Give the agent a Fractask parent and turn it loose.

“Here is the parentId for Acme onboarding. Use OpenClaw to advance the open parts. Log artefacts as you go. Stop when there is nothing left to claim.” That is the whole prompt.

04

Read the tree the next morning.

The CLI, the web app, your phone — all the same tree. You see what shipped, what stalled, and what is still claimed by the bot. Reclaim, edit, or move on.

Try it

Stop watching your agent type. Start reading what it did.

★ Star Fractask on GitHubVisit OpenClaw →

Fractask is MIT-licensed. OpenClaw lives at openclaw.ai. The integration is just two MCP entries in your client config.