CiTIUS · Universidade de Santiago de Compostela
Multi-agent AI,
built as graphs you can read.
A visual graph studio, experiments, datasets and labeling — one research platform for composing, running and measuring agent workflows.
Research preview · built and developed openly at CiTIUS
What it is
One place to compose, run and measure agents
Most agent work happens in scattered scripts: a prompt here, a retrieval step there, an evaluation notebook somewhere else. Chatty puts the whole pipeline on one canvas, where every step is a node you can inspect, rerun and compare.
Runs are first-class. A chat and an experiment trial are the same primitive underneath, so anything you can converse with, you can also sweep over a dataset and score.
Cost is tracked per run, per model, per node — because the interesting question about a panel of five agents is usually whether it was worth it.
One question, every model
Ask once. Compare side by side.
A session fans one prompt out to as many models as you like, keeps the answers side by side, and tracks what each one cost you.
You
Summarise the trial results and flag anything anomalous.
Score by cohort
Run summary
| metric | value |
|---|---|
| trials | 240 |
| mean score | 0.81 |
| total cost | $0.020 |
| flagged | cohort C |
Computational graphs
Agents as graphs you can read
Every workflow is a directed graph of typed nodes: models, tools, retrieval, branches, loops, whole panels of agents. Four the Studio runs today.
Chat with tools
A model that can reach for a tool mid-answer, then hand the result back into the conversation.
Retrieval-augmented answers
Semantic search over your own context items, injected into the prompt before the model runs.
Multi-agent panel
Several agents deliberate in rounds, an orchestrator routes the debate, and the panel converges on an answer.
Evaluation loop
Run a graph over a dataset, score every prediction against its reference, and collect the metrics in a dashboard.
Rendered by the Studio's own canvas — the same node components you edit with, lighting up the way a real run does.
The platform
Four pillars
Chats
Multi-model conversations with tools, memory strategies and streaming, over whichever providers you have configured.
Studio
A visual canvas for computational graphs: LLM nodes, tools, conditionals, loops, sub-graphs and multi-agent panels — edited collaboratively.
Research
Experiments with parameter sweeps, evaluation metrics and cost tracking, plus a logbook to write up what you found.
Labeling
Datasets and annotation jobs with typed questions, multiple annotators and inter-annotator agreement.
Your credentials
Bring your own models
Provider API keys
OpenAI, Mistral, Gemini, HuggingFace and NVIDIA Build keys belong to you or to your team, encrypted at rest with AES-256-GCM. There is no shared server key, so nobody spends anybody else's quota.
Your own Ollama servers
Point a workspace at any Ollama instance you can reach — a lab GPU box, your laptop — test the connection from the interface, and its models show up in every picker.
That does mean a brand-new account starts with no models. Adding a key or a server in Settings is the first thing to do after signing in.