← Docs
Getting started

Your first experiment

From sign-in to a completed training run. Follow along step by step — the whole thing takes about ~8 minutes.

What you will do
  1. Sign in~30 seconds
  2. Create your first session~1 minute
  3. Wait for your pod~2 minutes
  4. Send your first message~30 seconds
  5. Run a training script~2 minutes
  6. Watch the experiment~1 minute
  7. Export your results~30 seconds
  8. Stop the session~30 seconds
Total: ~8 minutes · 8 steps · No credit card required
1/8~30 seconds

Sign in

What to do

  1. Go to run.meshia.io and enter your email.
  2. Check your inbox for a 6-digit code. Enter it on the verification screen.

What you will see

A verification code input. After entering the code, the dashboard loads with a "New session" button front and center.

What it means

No passwords to remember. Meshia uses magic codes so there is nothing to forget or leak. The code expires after 10 minutes.

TipThe code sometimes lands in spam the first time. Check there if you don't see it within 30 seconds.
Next: Create your first session →
2/8~1 minute

Create your first session

What to do

  1. Click "New session" on the dashboard.
  2. Select RTX 4090 (24 GB) from the GPU list. It is the best default for getting started.
  3. Name your session something descriptive, like "first-experiment".
  4. Click "Launch".

What you will see

A GPU picker showing available hardware with VRAM, price, and availability. After clicking Launch, a status bar appears showing "Provisioning...".

What it means

Meshia is spinning up a dedicated GPU pod just for you. No one else shares this machine. The RTX 4090 gives you 24 GB of VRAM, which handles models up to 13B parameters comfortably. The launch page shows the current variable GPU rate before you start.

TipTrial accounts start with $0 and get free compute credits. No credit card needed to follow this tutorial.
Heads upDon't close the browser tab during provisioning. The session will still launch, but you will need to find it again on the dashboard.
Next: Wait for your pod →
3/8~2 minutes

Wait for your pod

What to do

  1. Watch the status indicator progress through the stages.
  2. The workspace opens automatically when the pod is live.

What you will see

A status bar cycling through: queued, provisioning, starting, running. Then the workspace appears: file explorer on the left, canvas in the center, chat on the right.

What it means

Cold starts take 2-3 minutes because Meshia is pulling the container image, allocating the GPU, and booting the agent. Subsequent sessions with the same GPU type start in under 60 seconds because the image is cached.

TipWhile waiting, you can read the workspace layout. Left panel has files, terminal, and GPU stats. Right panel is where you talk to the agent.
Next: Send your first message →
4/8~30 seconds

Send your first message

What to do

  1. Click the chat input at the bottom of the right panel.
  2. Type: What GPU do I have? Show me the CUDA version and available VRAM.
  3. Press Enter.
What GPU do I have? Show me the CUDA version and available VRAM.

What you will see

The agent responds within a few seconds with your GPU model (NVIDIA RTX 4090), CUDA version, driver version, and available VRAM. It runs nvidia-smi under the hood and formats the output for you.

What it means

The agent has full access to the pod. It can run any shell command, read and write files, install packages, and execute Python. Asking about the GPU confirms the hardware is live and the agent is connected.

Next: Run a training script →
5/8~2 minutes

Run a training script

What to do

  1. In the chat, ask the agent to train a small model.
  2. Watch the terminal panel on the left as the agent sets up the environment and starts training.
Train a small image classifier on CIFAR-10. Use a simple CNN, 5 epochs. Show me the loss after each epoch.

What you will see

The agent installs PyTorch, downloads a small dataset, writes a training script, and starts the run. You see real-time terminal output with loss values printing every few steps. An experiment card appears in the left panel.

What it means

The agent chose a small model (a 2-layer CNN on CIFAR-10) that trains in about a minute. This is fast enough to see the full lifecycle: setup, training, metrics, and results. For real work you would describe your actual task and the agent would scale accordingly.

Heads upIf the agent asks clarifying questions, answer them. It might ask about batch size or learning rate. For this tutorial, telling it to use defaults is fine.
Next: Watch the experiment →
6/8~1 minute

Watch the experiment

What to do

  1. Click the "Experiments" tab in the left panel (or press Cmd+E).
  2. Click on your running experiment to see its detail view.

What you will see

A live loss curve updating in real time as each epoch completes. Below it, a GPU utilization chart showing compute usage. The config panel on the right shows hyperparameters the agent chose. When training finishes, the experiment status changes to Completed.

What it means

Meshia automatically instruments training code to log metrics, capture configs, and track GPU utilization. You did not have to set up Weights & Biases or write any logging code. The agent handles it.

TipAfter training completes, try asking the agent "What was the final accuracy?" It can read the experiment results and summarize them in plain language.
Next: Export your results →
7/8~30 seconds

Export your results

What to do

  1. Go back to the experiment detail view and click "Export".
  2. Choose what to download: the model checkpoint, the training config, or a full experiment bundle.

What you will see

A download dialog with three options. The checkpoint is a .pt file. The config is a JSON file with every hyperparameter. The bundle is a zip containing both, plus the training script and metric logs.

What it means

Everything the agent produced is yours to keep. Download the checkpoint to use in your own code, share the config to let someone reproduce your results, or grab the full bundle for your records.

TipYou can also ask the agent to push results to a GitHub repo if you have the integration connected. Say "Push the model and training script to my repo."
Next: Stop the session →
8/8~30 seconds

Stop the session

What to do

  1. Click the "Stop session" button in the top toolbar (or press Cmd+Shift+Q).
  2. Confirm when prompted.

What you will see

A confirmation dialog showing your session summary: total time, GPU hours used, compute cost, and number of experiments. After confirming, the status changes to "Stopped" and you return to the dashboard.

What it means

The pod is destroyed and the GPU is released. You are only charged for the time the pod was running, billed per minute. Your experiment data, metrics, and artifacts persist in Meshia. Agent memory from this session carries over to future sessions on Pro plans.

Heads upMake sure to export or save anything you need before stopping. The pod filesystem is wiped on stop. Experiment artifacts saved through Meshia persist, but files left on the pod do not.
You did it.

You signed in, launched a GPU, trained a model, watched the metrics, and exported your results. That is the core Meshia loop. From here, explore what else the agent can do.

Explore agent roles →Experiment tracking →Autonomous research →
← All docsYour first session →