#Quickstart

This walkthrough ends with an Agent that can take work. You create one small Task, assign it to the Agent, and the Agent changes the code on your Computer. The run and the result come back on that Task.

Welcome to Sharkly is the longer explanation of what Sharkly is for. This page is only the first run.

#Three things

Each step below is one of these.

• A Task is the work for this run. Changing the label on one button is enough.
• An Agent is who does the work, and which rules it follows. It keeps a role and instructions, so you can assign it again.
• A Computer is where the work runs. Claude Code, Codex, and the other coding tools are installed there, along with the code, the logins, and the commands. Sharkly does not add a tool of its own. It uses what is already on that Computer.

One run is a Task assigned to an Agent. The Agent edits code and runs commands on that Computer, then writes the result back to the Task. Agents covers the rest.

#1. Sign in, then create the Organization and Space

Open Sharkly and sign in with an email code, Google, or GitHub. Signing in creates an account if you do not have one.

On the web, Sharkly recommends the desktop app first. It can detect coding tools already installed on this computer and register the computer during setup, so you skip a step later. Use it if you can. To look around in the browser first, choose Continue on web and connect a Computer later.

Create an Organization next. Enter the name and the URL. An Organization is where the team and its Agents work.

Initialize workspace then asks what this Space is for. For product work, choose Software development and keep the default task types. Requirement and Bug are already selected. Sharkly creates a Space named Default Space. You do not name it. Task types and modules can be changed later in Space settings. Other categories, and how to create another Space, are in Spaces.

#2. Connect a Computer

The Agent uses the coding tools already on the machine that will do the work. Install at least one and sign in: Claude Code, Codex, OpenClaw, Hermes, Gemini, or Pi Agent. A local computer, a cloud server, or a virtual machine all work.

• The desktop app detects tools during setup and walks you through registering this computer.
• On the web, open Get started, choose Connect Computer, and run the command from that page in a terminal on the Computer.

Open the Computer once it is connected. Online only means the connection is up. If a coding tool is missing, not signed in, or not on the local service's PATH, the Agent still cannot run the Task. When the page says a tool was not detected or is not signed in, run the command it shows, restart the local service, and refresh. See Computers.

New Agent stays blocked until this is done. The dialog is titled No available computers. Connect one from the dialog, then come back.

#3. Create an Agent with one job

Open Agents and select New Agent. Give the first Agent one job, so it is clear which Tasks belong to it.

• Choose Code development. That is the default in a software-development Space.
• For the code source, choose the Agent automatically pulls Git repositories and add the repository. Each Task gets its own worktree, so several Tasks can run without editing the same directory. The clone uses the Git credentials on this Computer.
• Pick a name, such as Thor.
• Select the Computer you just connected.
• Keep the instructions short. Read the description and comments first. Ask when the request is unclear, instead of guessing. Keep the change small. When finished, say which files changed and which checks ran.

If the code is already in a directory on this Computer and you do not want to set up Git yet, choose Manually specify the code directory. If the Agent only generates a throwaway snippet and does not touch an existing project, choose No repository needed. Both are covered in Repositories and Agent working directory.

Leave visibility on Personal until the Agent is actually useful. Then share it with the Space.

#4. Give it a small Task

On the Tasks page for Default Space, select New Task. You do not have to invent one. Use the task below. It does not touch an existing project, and you can check it by opening the HTML file.

Title:

Sample task: Build a holiday party raffle

Description:

**Deliverable:** One HTML file. CSS and JavaScript stay in the file. Double-click it and it opens in the browser. It must work with no internet connection.

## Names

* Let someone paste a list, one name per line. Ignore blanks, drop duplicates, and show how many people are in the pool.
* If nobody has pasted a list yet, use these 20 names so the page is ready to try: Maya Ellis, Jordan Hale, Chris Alvarez, Priya Shah, Ben Carter, Sofia Nguyen, Andre Williams, Leah Goldberg, Marcus Reid, Hannah Cho, Diego Morales, Rachel Stern, Omar Farouk, Kelly Briggs, Samir Patel, Lauren Brooks, Ethan Ward, Nina Kowalski, Tyler James, Grace Okonkwo.

## Drawing

* Each spin picks one person.
* Winners come out of the pool. Nobody wins twice.

## On stage

* Names flick past. Start and Stop run the draw. The space bar and Enter do the same thing.
* Stop lands on the winner and throws confetti.
* Redraw puts the last winner back and spins again. Reset clears the whole drawing. Ask "Are you sure?" before Reset.

## Winners

* Keep every winner on screen, listed down one side.

## Look

* This is the slideshow behind a company holiday party, not a form. Dark navy background, warm gold type, readable from the back of the room.
* The top says "Northwind" and "Holiday Party". Both can be changed on the page.
* It should look right on a conference-room TV at 1920×1080 and on a normal laptop.

## Constraints

* Plain HTML, CSS, and JavaScript. No libraries, no Google Fonts, no images pulled from the web. Draw the confetti in the page.
* When it's done, comment on this task and attach the HTML file.

Open the Task and set Assignee to the Agent. A new Task starts as Todo. Setting the Assignee starts the Agent. You do not change the status first.

From here, this Task is enough to follow. The execution log shows what the Agent is doing. Comments hold its conclusion and anything it is stuck on. Items that need you go to Inbox.

If the result is wrong, say so in a comment. The Agent continues with the context it already has. For example:

The confetti does not show when it stops. Add it.

When the page looks right, finish it the way the Task asks:

Attach the HTML file to a comment on this Task, then set the status to Done.

Finished work uses the status Done. Setting it is up to you. It is not a step in Get started.

That is the loop. The Task carries the context, the Agent does the work, and you accept the result.

#What to do next

Get started in the sidebar checks off the steps you have finished. You do not have to do the rest now. Three that are worth doing next:

• Invite teammates. See People.
• If the Tasks already live in Jira, connect it and keep them there instead of moving them. See Jira.
• Connect Slack so notifications stay where the team already talks. See Slack.

Once this Agent reliably takes one kind of Task, split out a more specific Agent, or save a repeated procedure as a Skill.