Supernova - All your data in Claude and Codex
by•
Supernova connects your startup’s live data to Claude and Codex, so anyone can ask questions, investigate performance, and run complex analysis in the AI tools they already use. Connect Stripe, HubSpot, PostgreSQL, and 30+ other apps, then analyze revenue, pipeline, customers, usage, and operations without waiting on engineers or moving everything into a traditional BI stack.
Replies
Love that Supernova meets teams inside Claude and Codex instead of forcing everyone into yet another BI dashboard, that alone removes so much friction from getting answers.
Supernova
@ilko_kacharov Thanks Ilko! Yeah, we all practically live in Claude and/or Codex these days, so it feels like the natural next step.
1752vc Pitch Deck Analyzer
Good to see you back for a second launch. Table permissions went from private preview to shipped for everyone while the thread was still running. Congrats on that turnaround.
Supernova
@lucasjpols Great to be back. PH is so great for product feedback.
Promomix
The concept is compelling but I’d want to know how you prevent confident nonsense when the underlying data is incomplete or inconsistent across systems.
Supernova
@margret_rhyme In those situations it can be a good idea to use the most capable AI model. Fable (or even Sol 5.6 on Extra High) are much more careful when it comes to comparing data from different systems and will often point those issues out themselves in fact.
Partly this is because we expose several things that help the AIs notice issues - each table and each row is tagged with its "freshness", so it's easy to see which tables have not been updated in a while, for instance. We expose whether rows were deleted which helps with accidentally deleted data.
When it comes to consistency, where AI models shine is when they can easily verify their own assumptions. The way Claude and Codex pull data from Supernova, they have a lot of data points to use to verify their own answers and check for inconsistencies. A join that doesn't match every row, or numbers that don't match up in two sources, can be spotted and the agents will point it out.
For a more concrete example, If the data in Stripe gives $182k but the data in Salesforce says $175k, the datasets themselves give Claude/GPT a starting off point to explore why it happened. If the underlying issue is that Salesforce is missing 6 contracts, this is easy to for the AIs to spot.
the thing i would want to know most: when someone asks what revenue looked like last month and the model has the right rows in front of it, how often does the number come back right?
we pointed eight models at a live pricing api recently and two of them misread a quantity ladder they had been handed correctly. not hallucination, the data was in context, they just read the wrong row. one was out by 4x, the other by about 6 percent, and the 6 percent one is the dangerous one because nobody double checks a number that looks plausible.
for a support reply that is an annoyed customer. for revenue analysis it is a number that ends up in a board deck. do you verify the arithmetic anywhere before it renders, or is that left to the model?
Supernova
@jernej_jan_kocica for important numbers (such as ones that end up on a deck), I run adversarial review agents, just like code. E.g. give something like this to Fable:
"This is a revenue analysis we don't fully trust. Do an adversarial review of the numbers, methodology and data sources used. These numbers will be used on a board deck so it is critical that we avoid mistakes."
Or maybe "Spawn multiple agents to cover distinct ways in which the numbers could be wrong"
In our own harness in-app we built this Orchestrator -> Implementer -> Reviewer workflow for that reason. For Claude unfortunately the MCP rules don't allow us to instruct the model on its behavior too strongly, but you can implement a similar workflow yourself.
We should probably publish some docs on this!
Of course, for something like an actual board deck I would still probably review manually myself.
@lukezapart that is more concrete than i expected, thanks. the orchestrator to implementer to reviewer split is the part i want to steal.
the bit i am still unsure about is whether an adversarial pass catches the quiet error. a reviewer asked to find problems with 39.00 finds it, because it is wrong on its face. a reviewer looking at 9.00 against the same price ladder has to read that ladder correctly in order to disagree, and something just misread it once already. plausible errors survive review for the same reason with people and with models.
your last line is the honest one though, and it is where we landed too. for anything that actually matters a person still reads it before it goes out. this is the third time this week someone has independently told me the answer is a separate verification pass, which is starting to make me think that is the state of the art rather than a workaround.
would genuinely read those docs if you write them.
We ship a remote MCP server too, and the thing that surprised me most was how misleading the tool-call metrics are. Ours read close to a 100 percent failure rate for a while. When I broke it down, 36 of 80 recorded tool calls were unauthenticated probes getting a 401, and 27 more named tools we do not publish at all. Seven were real calls from real clients, and all seven were the same bug.
So the number that looked like a broken product was mostly the open internet knocking on the door.
With 30 plus connectors exposed, do you separate authenticated traffic from probes before computing anything? And do you pass upstream errors to the model verbatim or normalise them? We were turning a 402 into a 5xx and it made the real failure unreadable for months.
Lancepilot
Meridian
Supernova’s permission controls and background syncing look promising, especially for large datasets. I’m curious how teams monitor sync failures or schema changes across integrations, and where the process still needs manual intervention.
How do you give Claude or Codex access to live company data without creating a new governance and maintenance burden?