Start with one AI answer. Escalate to a configurable model panel when another perspective is worth the time and cost. Stream the discussion and inspect the synthesis. MIT code.
1/4 · Not every question needs a panel of models. Sometimes one answer is enough, and sometimes it deserves a proper challenge.
Mercury starts with a single answer. When another perspective is worth the time and cost, you escalate to a configurable model panel, stream the discussion and inspect the synthesis. More scrutiny becomes a deliberate choice, not a quiet multiplication of calls.
Streaming the panel's discussion lets you see how a conclusion was reached, rather than only receiving the synthesis. That makes it easier to judge whether the added scrutiny was worth it.
I built it because the AI bill needs an owner, and so does the decision to trust an output. More models can offer different perspectives, but agreement between them isn't verification. You still need evidence, testing or human judgement suited to the task.
This is Day 6 of 11 of the Product Hunt Marathon: 11 free, open-source launches, one a day, 10-20 November.
The code is MIT-licensed, so you can adapt how and when escalation happens.
Give it a question you'd genuinely escalate. Did the panel change your thinking, or just make the answer longer?
2/4 · How escalation works and what it can't prove
If another perspective is worth it, you can escalate to a configurable panel using Quick debate, a saved workflow or a custom setup. I wanted access to different model perspectives without making every question go through the same elaborate process.
The distinction is important. More model calls are not automatically a better answer, and agreement is not verification. Check evidence separately and keep sensitive inputs within a provider boundary you are comfortable with.
My tribute to a questioning culture is not another machine that declares itself right. I want the reasoning and disagreement available for inspection. Shared evidence is more useful than a polished consensus nobody can challenge.
After yesterday's contract-review tool, today's launch, Day 6 of 11 of the Product Hunt Marathon, broadens the question: how do we decide when another view is worth the time and cost? Tomorrow, Day 7 of 11, it's Megamesh.
Mercury makes the budget side of CHO concrete. One accountable lead or small team can evaluate models, set spending limits and maintain shared workflows, instead of every colleague choosing tools and repeating experiments alone. That does not mean restricting all AI use to one person. It means giving outcomes and costs an owner.
My “SaaS is dead.” thesis would be hollow if an owned workflow simply created an uncontrolled inference bill. Start with a simpler route, escalate for a defined gap and count retries, review and upkeep. These are operating principles to evaluate, not savings already proved by Mercury.
I'd like examples where escalation reveals a useful weakness, and examples where it adds nothing. Both can improve how the workflow is designed. Please describe what a better answer would have enabled you to decide.
Mercury's MIT code is available to adapt or commercialise while keeping the notices. A good AI-assisted contribution could pair one focused change with a small evaluation set and an explicit stopping condition. Set the budget, inspect the implementation and test whether the change helps rather than merely lengthening the exchange.
I will prioritise findings that make the system easier to evaluate and use deliberately.
The Marathon is dedicated to the builders' mentality: making small, useful things in public, every day. Web Summit Lisbon is over now, and thank you to everyone I met there. I live in Lisbon and am happy to meet around the city; X (@paulfxyz), LinkedIn or paulfleury.com/contact will find me.
Separately from the Marathon, I'm also working on a few other ventures. Openline (https://openline.com) is my main venture and startup, where the aim is to reshape the eSIM market, with a deep pipeline of ideas ready to execute; I'm looking for VCs, angels and investors. Iris (https://iris.online) and Digital Spot (https://digital.spot) are spin-offs of tools built for Openline, and I'm looking for investors and early customers for both.
Enki (https://enki.ngo) is rather a thought experiment: a non-profit, local-first AI public good, too big for me alone, and I'm looking for philanthropists, donors and helpers. Cup of News (https://cupof.news) is a Sunday-afternoon hobby that is soon turning into an app, and I'm looking for contributors and users.
CHO stands for Chief Hacking Officer, a playful name for a serious role. I'm open to a CHO brief of up to 30h/week with a great company or startup, a worthwhile product, trust and a modest budget.
Replies
1/4 · Not every question needs a panel of models. Sometimes one answer is enough, and sometimes it deserves a proper challenge.
Mercury starts with a single answer. When another perspective is worth the time and cost, you escalate to a configurable model panel, stream the discussion and inspect the synthesis. More scrutiny becomes a deliberate choice, not a quiet multiplication of calls.
Streaming the panel's discussion lets you see how a conclusion was reached, rather than only receiving the synthesis. That makes it easier to judge whether the added scrutiny was worth it.
I built it because the AI bill needs an owner, and so does the decision to trust an output. More models can offer different perspectives, but agreement between them isn't verification. You still need evidence, testing or human judgement suited to the task.
This is Day 6 of 11 of the Product Hunt Marathon: 11 free, open-source launches, one a day, 10-20 November.
The code is MIT-licensed, so you can adapt how and when escalation happens.
Give it a question you'd genuinely escalate. Did the panel change your thinking, or just make the answer longer?
2/4 · How escalation works and what it can't prove
If another perspective is worth it, you can escalate to a configurable panel using Quick debate, a saved workflow or a custom setup. I wanted access to different model perspectives without making every question go through the same elaborate process.
The distinction is important. More model calls are not automatically a better answer, and agreement is not verification. Check evidence separately and keep sensitive inputs within a provider boundary you are comfortable with.
Try Mercury: https://mercury.sh
Repository: https://github.com/paulfxyz/mercury
The launch sequence: https://ph.paulfleury.com/
3/4 · Give the AI bill an owner
My tribute to a questioning culture is not another machine that declares itself right. I want the reasoning and disagreement available for inspection. Shared evidence is more useful than a polished consensus nobody can challenge.
After yesterday's contract-review tool, today's launch, Day 6 of 11 of the Product Hunt Marathon, broadens the question: how do we decide when another view is worth the time and cost? Tomorrow, Day 7 of 11, it's Megamesh.
Mercury makes the budget side of CHO concrete. One accountable lead or small team can evaluate models, set spending limits and maintain shared workflows, instead of every colleague choosing tools and repeating experiments alone. That does not mean restricting all AI use to one person. It means giving outcomes and costs an owner.
My “SaaS is dead.” thesis would be hollow if an owned workflow simply created an uncontrolled inference bill. Start with a simpler route, escalate for a defined gap and count retries, review and upkeep. These are operating principles to evaluate, not savings already proved by Mercury.
4/4 · What I'd like to learn from your tests
I'd like examples where escalation reveals a useful weakness, and examples where it adds nothing. Both can improve how the workflow is designed. Please describe what a better answer would have enabled you to decide.
Mercury's MIT code is available to adapt or commercialise while keeping the notices. A good AI-assisted contribution could pair one focused change with a small evaluation set and an explicit stopping condition. Set the budget, inspect the implementation and test whether the change helps rather than merely lengthening the exchange.
I will prioritise findings that make the system easier to evaluate and use deliberately.
The Marathon is dedicated to the builders' mentality: making small, useful things in public, every day. Web Summit Lisbon is over now, and thank you to everyone I met there. I live in Lisbon and am happy to meet around the city; X (@paulfxyz), LinkedIn or paulfleury.com/contact will find me.
Separately from the Marathon, I'm also working on a few other ventures. Openline (https://openline.com) is my main venture and startup, where the aim is to reshape the eSIM market, with a deep pipeline of ideas ready to execute; I'm looking for VCs, angels and investors. Iris (https://iris.online) and Digital Spot (https://digital.spot) are spin-offs of tools built for Openline, and I'm looking for investors and early customers for both.
Enki (https://enki.ngo) is rather a thought experiment: a non-profit, local-first AI public good, too big for me alone, and I'm looking for philanthropists, donors and helpers. Cup of News (https://cupof.news) is a Sunday-afternoon hobby that is soon turning into an app, and I'm looking for contributors and users.
CHO stands for Chief Hacking Officer, a playful name for a serious role. I'm open to a CHO brief of up to 30h/week with a great company or startup, a worthwhile product, trust and a modest budget.