Connect your machine learning research stack

Weights & Biases, GitHub, OpenReview, and Google Calendar. Each can be connected through the settings page, and each is optional.

All four can be connected by opening Settings from your profile menu and going to Integrations. Each can be disconnected later from the same panel.

Weights & Biases

Give Althea a model architecture and a dataset, and it writes the training script, wires in W&B tracking, and submits the job to your remote server (opens in new tab). Training jobs run as background tasks, so the session sleeps while the model trains. When the run is done, the logs are in W&B to read together and decide the next iteration.

This is the grad-student loop, minus the grad student staying up to babysit the run.

You

Altheawrites the script, submits the job

Training jobyour remote server

Weights & Biasesmetrics · checkpoints · artifacts

You

Altheawrites and submits

Training jobyour remote server

Weights & Biasesmetrics · checkpoints

The W&B loop. The quiet stretch after the job's last metric is the session asleep.

The connection takes one key. Get API Key on Wandb opens the W&B authorization page; paste the key into Althea and click Connect. Althea verifies the key against the W&B API before storing it as a secret, and the panel then shows the connected W&B username.

When a code workflow needs experiment tracking, the runtime resolves the secret and exposes it to the job as WANDB_API_KEY. Remote-server jobs receive it too, forwarded by the submission helper. The training scripts Althea generates guard on it before initializing:

python

import os
import wandb
 
if os.environ.get("WANDB_API_KEY"):
    wandb.init(project="my-experiment", config={"learning_rate": 0.001})
    wandb.log({"loss": 0.42})

Asking for a small ResNet on CIFAR-10 with the training metrics logged to W&B works, and so do learning-rate sweeps with a run per configuration, or a remote-server run that saves checkpoints and uploads the final model artifact.

GitHub

GitHub gives the Code Agent repositories to work with. It can clone them, search them, and open pull requests, on public code or your own.

Connecting uses a personal access token. From the GitHub panel, Create Token on GitHub opens GitHub's token creation page with the required permissions pre-selected; give the token a descriptive name, set an expiration, generate it, and paste it back into Althea.

The GitHub integration panel in Althea's settings — The GitHub panel under Settings, then Integrations.

GitHub's token creation page with the required permissions pre-selected — GitHub's token page, opened with the required permissions already selected.

Once connected, a claim check runs against real code. Ask "Find the official code base of the FSNet paper (NeurIPS 2025), clone the repo and figure out if they use dropout" and the Code Agent finds the repository, clones it, and runs a search for "dropout" across the codebase. Within seconds you get a clear answer, with the relevant files and line numbers.

The ask

A claim about a published paper, checked against its code.

Researcher

Find the official code base of the FSNet paper (neurips 2025), clone the repo and figure out if they use dropout.

The search

Run by the Code Agent across the cloned repository.

Althea

Code AssistantInstructions: Clone the git repository from https://github.com/MOSSLab-MIT/FSNet and search the entire codebase for the term "dropout". Report back with the files and line numbers where "dropout" is mentioned, if any.

# Assistant: I'll help you clone the repository and search for the term "dropout" in the codebase.git clone https://github.com/MOSSLab-MIT/FSNetCloning into 'FSNet'...Grep: {"pattern":"dropout","path":"/home/user/work/FSNet","output_mode":"content","-i":true,"-n":true}FSNet/utils/trainer.py:100: dropout = config["dropout"]FSNet/utils/trainer.py:105: model = MLP(data.xdim, hidden_dim, out_dim, num_layers=num_layers, dropout=dropout)FSNet/utils/trainer.py:107: model = MLP(data.xdim, hidden_dim, data.ydim, num_layers=num_layers, dropout=dropout)FSNet/utils/trainer.py:334: # Dropout decayFSNet/utils/trainer.py:337: if isinstance(m, nn.Dropout):FSNet/utils/trainer.py:341: if isinstance(m, nn.Dropout):FSNet/models/neural_networks.py:4:# Multi-Layer Perceptron (MLP) with adjustable dropout and layer countFSNet/models/neural_networks.py:7: def __init__(self, input_dim, hidden_dim, output_dim, num_layers=1, dropout=0.0):FSNet/models/neural_networks.py:13: if dropout > 0:FSNet/models/neural_networks.py:14: layers.append(nn.Dropout(p=dropout/(i+1))) # Adjust dropout rate for each layerFSNet/main.py:35: parser.add_argument('--dropout', type=float, help='Dropout rate for the model')

The answer

The repository confirmed and linked.

Althea

I've located the official codebase for the FSNet paper from NeurIPS 2025 and can confirm that it does use dropout.

The code is available at the following GitHub repository: https://github.com/MOSSLab-MIT/FSNet

My analysis of the codebase shows that "dropout" is a configurable parameter for the neural network models, and it's mentioned in several files, including the model definition, training scripts, and configuration files.

Let me know if you have any other questions!

The same works on your own repositories. Ask a question about your own codebase in plain language and the answer comes back the same way. And if W&B is connected too, Althea can combine them and push experiment code to a private repository while the runs log to W&B.

OpenReview

Connecting OpenReview attaches a verified research identity to your account. The network uses it when provenance about people matters. It routes invitations, checks membership in private groups, and matches researchers across institutions. Verified identity covers what it checks and why it is separate from your researcher profile; for ordinary chat, literature search, or code execution, you can leave it disconnected.

Setup is one authorization: click Connect OpenReview in the panel, complete the flow on OpenReview, and the panel confirms the connection when you return.

Google Calendar

Calendar has two access modes, and they connect and revoke separately. Each mode is its own button in the Calendar panel: Connect View-Only or Connect Assistant. Althea redirects you to Google to authorize just the scope you picked, and when you return, the panel shows that mode as connected.

Mode	What it grants	Try saying
Briefings (view-only)	Reads events, modifies nothing. For schedule awareness, meeting preparation, and briefings	"What meetings do I have tomorrow?" · "Brief me on my next three meetings" · "Do I have any open time on Friday afternoon?"
Full Assistant	Create, update, and delete, at your direction. For coordinating follow-ups, moving events, setting reminders	"Schedule a 30-minute follow-up with Nasim tomorrow afternoon" · "Move my 3pm meeting to Thursday if there is an open slot" · "Create an all-day event called Vacation next Monday"

View-only requests stay in the conversation. Althea retrieves the events and summarizes what is relevant, whether that is tomorrow's schedule or a specific timeframe. Scheduling resolves dates and times against your browser timezone and the one configured in your Google Calendar.

Disconnecting

Every connection is removed from the panel it was made in. Disconnecting W&B revokes the stored secret, so future code-agent and remote-server jobs stop receiving the key. Disconnecting OpenReview removes the identity connection and leaves Memory and the profile in place. The two Calendar modes are independent; revoking View-Only does not revoke Full Assistant, so each disconnects separately.