We will be undergoing planned maintenance on Oct 7th 6:00AM UTC / Oct 7th 2:00AM ET

TaskPilot

The all-in-one AI browser agent that helps you lock in, stay focused, and get real work done. TaskPilot turns goals into manageable steps, organizes your tabs, and guides you when you’re stuck.

Comment

Inspiration

We built TaskPilot because the browser is where focus goes to disappear. You open one tab to work, then end up with dozens of tabs, unfinished tasks, distractions, and no clear idea where to begin.

We wanted to create an agent that understands the user’s goal, helps them take the next step, and remembers their work across sessions.

What it does

TaskPilot is an all-in-one AI browser agent that helps users lock in and get work done.

Users describe a goal, and TaskPilot:

  • Breaks it into manageable steps.
  • Helps prioritize what to do next.
  • Guides users through unfamiliar webpages.
  • Highlights the next button or field to interact with.
  • Detects when the user actually completes the action.
  • Organizes tabs around intent instead of website domain.
  • Detects distractions without blocking the web.
  • Saves and restores task sessions.
  • Gathers context from Gmail, Google Calendar, and GitHub.
  • Prepares approved actions such as GitHub issues, Gmail drafts, and calendar follow-ups.

TaskPilot never silently sends messages, creates records, or submits forms. Users review and approve every external action.

How we built it

TaskPilot is a Chrome extension built with TypeScript, React, and Manifest V3.

The extension includes:

  • A task planner.
  • A DOM-aware website navigator.
  • An AI tab organizer.
  • A focus and distraction agent.
  • Persistent task sessions.
  • A contextual Actions tab.

Our backend runs on Cloudflare Workers. Workers orchestrate the AI agents and connected-app workflows, while Durable Objects persist sessions, action previews, approvals, and resume state for each user.

We use structured model outputs so plans, navigation decisions, classifications, and action previews are validated before reaching the UI.

Composio connects TaskPilot to tools such as Gmail, Google Calendar, and GitHub. Read actions gather context, while write actions require explicit approval.

Challenges we ran into

The hardest challenge was making the agent useful without making it dangerous.

We had to ensure that TaskPilot could understand webpages without sending private page contents or form values to the model. We also needed to prevent it from filling sensitive fields or taking irreversible actions automatically.

Another challenge was handling messy real-world context. Browser tabs, email, calendars, and GitHub issues often contain duplicated, stale, incomplete, or conflicting information. TaskPilot identifies uncertainty and asks for clarification instead of confidently making something up.

We also had to make guide progression reliable. TaskPilot waits for the user’s real click before marking a step complete, preventing repeated instructions and incorrect progress.

Accomplishments that we’re proud of

We are proud that TaskPilot turns a vague goal into an active browser workspace.

The agent can guide users through webpage interactions, organize tabs by task intent, detect distractions, and restore work sessions later.

We are also proud of the safety model. Every external write is previewed, editable, and approval-gated. Credentials remain outside the extension, and sensitive browser data is excluded from model requests.

Most importantly, TaskPilot combines browser guidance, persistent memory, focus management, and real-world actions in one workflow.

What we learned

We learned that a useful agent needs more than intelligence. It needs state, context, clear boundaries, and the ability to recover when information is incomplete.

We also learned that human approval is a feature, not a limitation. Users want automation, but they also want to understand what will happen before an action is taken.

Finally, we learned that Cloudflare Durable Objects are a strong fit for agent workflows because they let TaskPilot maintain persistent, user-specific state while Workers coordinate tools and reasoning.

What’s next for TaskPilot

Next, we want to make TaskPilot more personalized and proactive.

We plan to improve:

  • Personalized profiles and working preferences.
  • Smarter prioritization based on deadlines and energy levels.
  • Better cross-source conflict resolution.
  • More contextual Gmail, GitHub, and calendar workflows.
  • Shared project workspaces for teams.
  • Stronger task recovery after interruptions.
  • More integrations through Composio.

Our long-term vision is for TaskPilot to become the layer between a person’s intentions and the tools they use every day.

Built With

Share this project:

Updates

Submission history