Start with something you already need to do.

What Hydra is

When you first open Hydra, the sidebar can make it look like a dashboard for other tools. But this is a real terminal. Here I am running an ordinary Python command, just as I would in my usual terminal.

Hydra brings together a Rust terminal, tmux-style session management, and remote access. The sidebar is React. The terminal uses a native GPU renderer. In this guide, we will set up one local project from the beginning.

Create your project

I have already installed Hydra and OpenCode. Your agent's installation and sign-in are separate. Start with a coding agent that already works on your machine.

Let's create a workspace for a small research task. This example uses synthetic sensor readings, not private research data. Click New project at the bottom of the sidebar.

I will call this Sensor Study. Use a name that tells you what the work is, rather than which model happens to be helping with it.

Next, enter the folder containing your project. This is an existing folder with a Python script, a dataset, and two tests.

Find previous sessions

Earlier, I discussed this dataset with OpenCode outside Hydra. I want to pick up that conversation, not start over.

Choose OpenCode, and Hydra finds its previous conversations for this folder. Here is Sensor Study planning. These are real saved conversations, including ones that began before I created this Hydra project.

I select the planning conversation. Notice that the launch mode changes to resume. We are opening the selected conversation, not merely launching a new agent in the same folder. If you want a clean start instead, choose fresh.

You can review the launch choice before creating the project. I am leaving automatic permission bypass off for this example.

I choose a colour to make the project easy to recognise, then click Create. Hydra adds the project to the sidebar and starts the selected session.

Continue the conversation

Here is the conversation we started outside Hydra, including its original question and answer. We were checking how to handle missing readings before comparing the two sensors.

Let's continue that discussion. I ask why missing readings should not count as zero. I am keeping this first task read-only so we can learn the workspace without having an agent change the project.

The response stays in the same conversation. Bringing existing work into Hydra does not mean giving up the context you have already built.

Understand the workspace

There are three levels to the workspace. The project groups related work. A window holds a terminal layout. Each pane is an individual terminal, running an agent or a shell.

Using the window menu, I rename this one Analysis. The same name appears in its tab at the top of the terminal.

Add a shell for checks

Now I want a separate place to check the work. Click the window button and call this one Run and results. Keeping discussion and execution separate makes it easier to find both.

For this window, choose Terminal instead of a coding agent, then Create. A Hydra workspace can contain ordinary shells as well as agent sessions. You do not need an agent for every task.

Run and check the work

In this build, the new shell starts in a scratch folder. I explicitly change to our project folder and check it with pwd before running anything. Use the path to your own project here.

Now I run the Python script. Each sensor has three valid readings. The missing entries are excluded, just as we discussed with the agent. These are the actual results from our small example dataset.

I also run the two unit tests. Both pass. The agent's explanation is useful, but I still check the work myself.

Give the session a name

Click Analysis to return to the conversation. The results remain in their own window, ready when we need them again.

The pane can have its own name too. Open its menu and choose Rename, just as we did for the window.

I call this Dataset review. Now the project, window, and conversation each have a useful label. This matters more as you add work. You can find the task you need without opening every terminal to remember what it contains.

Make room and switch tasks

Once the work is organised, I can hide Hydra's sidebar using the arrow at the top. Its space goes back to the terminal. The panel that OpenCode displays on the right belongs to OpenCode, not Hydra.

The window tabs stay visible. Click Run and results to check the same shell and test output, even with the sidebar hidden.

Then click Analysis to return. Switching tabs does not start a replacement conversation. This is still our planning session, with the follow-up we asked earlier.

I tell the agent that the tests passed and ask what to check next. This is the everyday loop. Discuss the work, run it in a shell, inspect the result, and continue the conversation.

When I want the project tree again, I click the handle on the left. The names and windows are still there. Use the sidebar to organise, and the tabs to move between tasks.

A multiplexer built in

So far we have used one pane per window. But Hydra is also a terminal multiplexer. That applies to ordinary shells, not just coding agents.

In Run and results, I choose Split down and keep Terminal selected. I leave the working directory set to Window folder in this build.

Click Split, and now there are two ordinary shells in the same window. Hydra provides this tmux-like layout and session management itself. You do not need to run tmux inside it.

Three agents in one window

Let's use the same idea for our agents. I return to Analysis and choose Split right. Instead of putting every conversation in a different window, we can keep related work side by side.

I name the new pane Methods check and leave OpenCode selected. This time I choose a new session. It is a separate conversation, not another view of the Dataset review conversation.

I give this agent the summary of our synthetic data and ask what we should check before comparing the sensors. Giving each conversation a specific job makes the layout useful. This one reviews the method, while the original conversation keeps the dataset discussion. I am asking for advice, not file changes.

Next I use Split down and name the new pane Test review. I keep OpenCode selected and start another new session. The goal is to separate the test discussion from the dataset and method reviews, while keeping them in one window.

We now have three OpenCode conversations in one window. In Test review, I describe our existing tests and ask which edge cases to check next. Our plain shells remain in Run and results.

Here is its answer. Each agent keeps its own conversation, and I can compare their suggestions without changing windows.

Arrange the work around you

The first layout does not have to be the final layout. I use the upward arrow on Test review to swap it with Dataset review. Their positions change, and each keeps its own conversation.

Now I use Swallow on Dataset review. It extends across the bottom row, giving that conversation more width. The other two panes remain above it. I have changed the layout, not merged the conversations.

Close a pane and bring it back

Suppose I want Test review out of the way for a moment. I click its close button. It leaves the terminal layout, but notice that Test review is still listed in the sidebar, now dimmed.

To bring it back, I drag the dimmed session name from the sidebar into the right side of the terminal area and release. Test review returns with the same conversation and answer. I do not start a new agent.

Use the full terminal space

With three panes, hiding the sidebar makes an even bigger difference. Click the arrow, and that space goes back to the terminal. All three conversations remain visible.

The tabs still work. Run and results brings back our two shells. Analysis brings back the three agent panes in their arranged positions. Each window keeps its own layout, so moving between discussion and execution is just a click.

Click the handle on the left to bring the sidebar back. Organise with the sidebar, rearrange panes with the mouse, and hide the tree when you want the space for your work.

Try one project

Start with one project, an existing conversation, and a shell for checks. Add more panes when the work needs them. You do not have to learn every control on the first day.

You do not need to change your whole workflow on the first day. Try this loop with your own project, and tell us where you get stuck.