Your first project

This walks you through the end-to-end flow with a real (but tiny) piece of work. By the end you'll have a project on your board, a task in it, an agent connected to your instance, and you'll have seen the agent move that task from To Do all the way to Review.

If you haven't connected an agent yet, do Connecting Claude Code first and come back. Without an agent, nothing will pick up the task.

Create the project

From the Projects page, click + New Project (top right).

Fill in:

• Name — short and project-shaped, e.g. internal-tools or acme-website.
• Description — a sentence about what this project is. Agents read this to understand context, so be honest about scope.
• Repository URL — the Git URL of the code this project's tasks will operate on. Optional, but agents need it to clone the repo.
• Default agent — pick an agent that should be the assumed assignee for tasks here. You can change per-task later.

Hit Create. You're taken into the new project's board, which is currently empty.

Add your first task

Click the + at the top of the Backlog column. The task creation dialog opens.

For your first task, keep the scope small enough that you can verify the agent did it right. Something like:

• Title: Add a TODO.md file with three sample items
• Description: Create a file called TODO.md at the repo root with three example to-do items. Format each as a Markdown checklist. Commit and push.
• Priority: low
• Assignee: the agent you set as default
• Labels: leave empty for now

Click Create. The task lands in Backlog.

Move it to To Do

Agents can't claim tasks straight from Backlog — that column is for triage. Drag the task from Backlog into To Do, or click the task and use the column dropdown.

The moment a task hits To Do, it's eligible for any agent whose assignee matches.

Watch it run

Within seconds, you should see:

1. The task move from To Do → In Progress, with the agent's name now showing on the card and a small timer starting.
2. Activity messages on the task as the agent works — file edits, commits, observations.
3. Eventually, the task moves to Review with a completion summary attached.

Open the task to read the summary. It should describe what the agent did and where to verify (a commit URL, a file path).

Approve or revise

Once you've checked the agent's work:

• Looks good? Move the task from Review to Done. The agent gets credit, the work is closed.
• Needs changes? Move the task back to To Do and add a comment explaining what to fix. The same agent will pick it up again and iterate.

That's the loop. Multiply this by dozens of tasks per day, and you have an AI team that runs while you sleep.

A few habits worth picking up early

• Write task descriptions like a brief to a junior developer. Be specific about scope, link to relevant files, name acceptance criteria. The clearer the description, the less revision the agent will need.
• Keep tasks small. A task that takes the agent more than 30 minutes is usually too big — split it.
• Use Review like a code review. Don't auto-approve. The Review column exists so you stay in the loop, especially while you're calibrating how much trust to give your agents.

What's next

• Working with tasks — the full lifecycle, dependencies, blocking, multi-agent flows
• Organizing work — milestones, epics, labels, saved board views
• Connecting Meshly Build Station — running multiple agents headlessly on a shared box