People · Tech · Research · Impact

Building cures with AI agents.

A launchpad where tokens fund real‑world disease research using AI agents.

Launch a project See how it works
Scroll to wake the nurse

The thesis

Research. Developers. Community. All aligned to cure real diseases.

The problem

Medicine is not short of ideas. It is short of the money and the hours to test them. A new drug takes more than a decade to reach patients, most candidates fail on the way, and thousands of diseases are too rare for anyone to pay for the attempt.

What changed

Two things. AI agents can now do real research work: read the literature, propose hypotheses and run the analyses, without tiring. And a community can now fund one precise goal directly, with every dollar traceable.

What we do with it

Cure-Fi puts the two together. Every coin backs a disease and adds agents to the fleet working on it, so the more coins a disease has, the faster it moves. Agents do the heavy lifting, researchers steer, labs confirm, and everything they find is published.

Time

10.5 years

is how long a drug takes, on average, to get from its first human trial to approval.

Odds

7.9%

of drug programs that enter human trials end in approval.

Neglect

10,000+

rare diseases are known. About one in twenty has an approved treatment.

Sources: BIO, Informa Pharma Intelligence and QLS Advisors, Clinical Development Success Rates 2011–2020 (2021); NIH National Center for Advancing Translational Sciences (2024).

How it works

From a token to a treatment.

  1. 01

    Launch

    One disease, many coins

    Anyone can launch a coin for a disease on the list, and a disease can have many coins, all adding to the same fleet. Launch without picking one, and the coin funds the shared pool, which backs the diseases nobody has picked yet.

    Who
    Anyone with a Solana wallet.
    Comes out
    A coin tied to one disease, or to the shared pool.
    On the record
    The coin, its disease and its fee split.
  2. 02

    Fuel

    Fees that hire agents

    Every trade pays a creator fee. The launcher chooses how much of it goes to research, at least 40%, and the split is locked on-chain the moment the coin is created. The research share goes to the Cure-Fi treasury and pays for the agents' work.

    Who
    The launcher sets the split; every trade adds to it.
    Pays for
    The agents' compute and data. Nothing else.
    On the record
    Every fee, every split and what it bought.
  3. 03

    Swarm

    Every coin adds agents

    Every 0.2 SOL of research fees adds three agents to the disease's fleet. Teams take different angles, from approved drugs to new targets, and all of them read, propose, test and challenge at once, around the clock.

    Who
    Fleets of four kinds of agent, growing with every coin.
    Comes out
    Ranked hypotheses, with the evidence for and against.
    On the record
    Every run: what was read, tried and concluded.
  4. 04

    Verify

    People decide, labs confirm

    Agents can be wrong, so nothing leaves the screen on their say-so. Every result is attacked by an agent running another company's model, then researchers choose which leads deserve an experiment, and a lab has the last word.

    Who
    The Skeptic, reviewers and partner labs.
    Comes out
    Experimental results, positive or negative.
    On the record
    Which leads were chosen, why, and what the lab found.
  5. 05

    Breakthrough

    Every result compounds

    Results are published as soon as they hold up, with their data and code. The dead ends are published too. The next team, and the next disease, start from everything already learned, so the whole network gets smarter with each one.

    Who
    Everyone. The findings are open.
    Comes out
    Papers, datasets and code anyone can build on.
    On the record
    The findings, the data behind them and the failures.

Mission control

Science you can watch happen.

Search any disease. See which agents are on it, what they are testing and exactly where the funding goes.

Live

Research fees collected

…

3 agents per 0.2 SOL

Live projects
in the app

Agents funded

0

The workforce

Agents that never clock out.

Every funded disease gets a fleet of specialist agents, and every coin makes it bigger. Each agent does one job, hands its work to the next, and leaves a record of everything it did. They run around the clock. People decide what happens with what they find.

The loop Read Propose Test Challenge Then round again. A researcher signs off before anything goes to a lab.

Literature agent

The Reader

Watches the journals, preprint servers and trial registries for anything that touches its disease. It pulls out each paper's claims, methods and results and files them in a shared evidence map, every entry linked back to its source.

Takes inPapers, preprints, trial recordsHands onAn evidence map, every claim cited

Hypothesis agent

The Inventor

Searches the evidence map for gaps, contradictions and findings nobody has put side by side, and turns them into hypotheses that can be tested, each with the result that would prove it wrong.

Analysis agent

The Analyst

Does the computational work: statistics on open data, structure prediction, virtual screens. Every run is saved with its code, data and settings, so anyone can repeat it.

Review agent

The Skeptic

Works against the others on purpose. It re-runs the analysis, checks every claim against its source and hunts for the usual ways results fool people: leaked data, cherry-picked samples, hidden confounders. Only what survives goes forward.

Not science fiction

Crews like this already make discoveries.

Two agent systems were published in Nature in 2026. Both proposed treatments, and people tested them in the lab, where they held up. Neither is a medicine yet: that takes trials, and years.

The breakthrough engine

Thousands of agents. One cure.

Every coin adds agents, so one disease can have hundreds and the network thousands. One finds a clue; together they find the pattern. When their work clicks together, a breakthrough is born.

Capabilities

What the agents can actually do.

Thirteen jobs, from reading the literature to writing the experiment that tests it.

Jump to

    Each capability is built on open tools and public data. Standard means it is a well-trodden pipeline today; emerging means it works and the field is still arguing about how well; frontier means it has been demonstrated in papers but is not yet routine. Nothing here is running on Cure-Fi yet: this is the specification the platform is being built to, and it is published as capabilities.json so the backend can be built from the same list this page shows.

    Built on

    Under the hood

    Not a chatbot. A research engine.

    A chatbot answers a question and forgets it. An agent is handed a goal, tools and a memory, and keeps working until the job is done or it can say exactly why not. This is how ours are designed.

    Anatomy of an agent

    Four parts, one loop.

    Every agent on Cure-Fi is built the same way. The loop at its heart is what makes it an agent rather than a chatbot.

    01 · A model that reasons

    It plans before it acts.

    A large language model reads, plans the next step and judges what came back. Models are improving quickly (see the trajectory below), so it is built to be swapped for a better one.

    02 · Tools that act

    It does, not just says.

    Literature and database search, a sandbox for running code, structure and property predictors. Every tool call is recorded.

    03 · A memory that lasts

    It starts from what is known.

    The evidence map holds every claim an agent has extracted, pinned to its source, so no agent starts from zero.

    04 · A loop that checks itself

    It knows when it is done.

    Plan, act, check, repeat. A run ends with a result and its full trail, or with the reason it stopped.

    What they read

    Much of biomedicine's raw material is already public. No person can read it all. An agent can read far more of it.

    The literature

    41,278,872

    citations and abstracts indexed in PubMed

    1,882,034

    of them dated 2025 alone: more than 5,000 a day. No person can keep up.

    PubMed

    Proteins

    150,006,383

    protein sequences cataloged

    UniProtKB

    Compounds

    2,921,148

    compounds, with 24.5 million lab measurements of their activity

    ChEMBL 37

    Counts read from each database's public interface on 14 October 2026.

    Five gates between an idea and a claim

    Agents can be confidently wrong. The process is designed on the assumption that they will be: every finding has to pass five gates.

    1. Cited

      No source, no claim. Every statement in the evidence map links to the paper, trial record or dataset it came from.

    2. Reproducible

      Every analysis is saved with its code, data and settings, so anyone can run it again and compare.

    3. Attacked

      A separate agent has one job: break the result. Re-run it, check the sources, hunt for the usual ways data misleads.

    4. Signed off

      A named researcher reviews what survives and decides whether it has earned an experiment.

    5. Tested for real

      The lab has the last word. Until it has spoken, a finding is a lead, not a fact.

    The evidence

    What the past has already proven.

    Cure-Fi rests on three bets: that AI can do real science, that a community can fund a medicine, and that open results compound. Each already has a track record. Here it is, marked with how far each result actually got.

    01AI can do real science

    Four results, oldest first. The bar under each shows how far it has travelled, from a computer's prediction to an approved medicine.

    2020AI-identified

    Halicin: a new antibiotic candidate

    A neural network trained to recognize antibacterial molecules picked out halicin, a compound structurally unlike conventional antibiotics. In the lab it killed a wide range of bacteria, drug-resistant strains among them, and in mice it treated a pan-resistant infection.

    1. Computer
    2. Lab
    3. Animals
    4. Patients
    5. Approved

    Stokes et al., Cell 180, 688–702 (2020)

    2020 → 2022AI-identified · repurposed

    Baricitinib: an approved drug, a new use

    In February 2020 researchers at BenevolentAI and Imperial College London used a knowledge graph, which includes connections extracted from the scientific literature by machine learning, to search approved drugs for one that might help against the new coronavirus. It identified baricitinib. In May 2022 the FDA approved it for adults in hospital with COVID-19 who need breathing support.

    1. Computer
    2. Lab
    3. Animals
    4. Patients
    5. Approved

    Richardson et al., The Lancet 395, e30–e31 (2020) · FDA approval letter, NDA 207924/S-006 (2022)

    2025AI-designed

    Rentosertib: designed by AI, tested in patients

    The target was identified by AI and the molecule generated by AI, in roughly 18 months from target to preclinical candidate. In a 71-patient trial, lung capacity rose by 98.4 ml on average over 12 weeks on the highest dose, and fell by 20.3 ml on placebo. It was a small, company-sponsored trial whose main goal was safety. A Phase 3 trial planned for 320 patients began recruiting in September 2026.

    −1000+100+200+98.460 mg−20.3Placebo

    Change in lung capacity after 12 weeks, ml. Mean and 95% interval. Each bar is one group’s own change, not a test of the difference between them.

    1. Computer
    2. Lab
    3. Animals
    4. Patients
    5. Approved

    Xu et al., Nature Medicine 31, 2602–2610 (2025) · Ren et al., Nature Biotechnology 43, 63–75 (2025) · ClinicalTrials.gov, NCT07687459

    2026Agent crews

    Co-Scientist and Robin: agents propose, labs confirm

    Two multi-agent systems proposed treatments from the literature that held up when people tested them in the lab. Robin also analyzed the experimental data: its paper states that every hypothesis, experimental direction, data analysis and data figure in the main text was produced by the system.

    1. Computer
    2. Lab
    3. Animals
    4. Patients
    5. Approved

    Gottweis et al., Nature 655, 487–496 (2026) · Ghareeb et al., Nature 655, 497–505 (2026)

    02A community can fund a medicine

    A patient charity, aiming its money at one disease, has already helped produce medicines.

    From the first investment to Trikafta took nineteen years. The model worked; the pace is what agents are built to change.

    Cystic Fibrosis Foundation, Our Venture Philanthropy Model (archived copy) · FDA approval package, Kalydeco, NDA 203188 (2012) · Cystic Fibrosis Foundation, press release, November 2014 (archived copy) · FDA approval letter, Trikafta, NDA 212273 (2019) · Cystic Fibrosis Foundation, Understanding Changes in Life Expectancy (archived copy)

    1. 1998Median predicted survival with cystic fibrosis is 32 years.
    2. 2000The Cystic Fibrosis Foundation, a patient charity, puts $40 million into Aurora Biosciences (now Vertex Pharmaceuticals), to hunt for compounds that correct the faulty protein.
    3. 2012The FDA approves Kalydeco, for patients with one specific mutation.
    4. 2014The Foundation sells its royalty rights for $3.3 billion, money it can put back into its mission.
    5. 2019The FDA approves Trikafta, for patients with at least one copy of the F508del mutation.
    6. 2025People with cystic fibrosis born between 2021 and 2025 are predicted to live to a median age of 66. A median for a group, not a promise to anyone.

    03Open results compound

    One open result can move a whole field.

    In 2020 Demis Hassabis and John Jumper presented AlphaFold2, an AI model that has since predicted the structure of virtually every protein known to science. Its predictions are open to every scientist, free. The databases counted above are free to use in the same way.

    The Nobel Prize in Chemistry 2024, press release · AlphaFold Protein Structure Database, EMBL-EBI

    • 200 millionprotein structures predicted
    • 2 million+people in 190 countries have used it
    • 2024Nobel Prize in Chemistry

    What is not proven yet

    • No medicine designed by AI, and none discovered by an AI agent, has been approved by the FDA yet, as far as we can find. The results above are lab findings, early trials and one repurposed drug.

    • AI has not yet beaten the hardest stage of trials. Molecules discovered with AI pass Phase 1 far more often than usual, 80 to 90% of the time. In Phase 2 they succeed about 40% of the time, on a small sample: about the industry's historic average.

      Jayatunga et al., Drug Discovery Today 29, 104009 (2024)

    • Helping in a clinic is harder than passing a test. In a 2026 randomized trial across 16 primary care clinics, giving clinicians an AI assistant did not significantly reduce treatment failures.

      Agweyu et al., Nature Medicine 32, 3032–3039 (2026)

    That is why Cure-Fi funds research with human review and lab confirmation built in, and promises no cure.

    The trajectory

    And the agents keep getting better.

    Research takes years. If the trend holds, the agents that finish a project will be far more capable than the ones that started it. Three measurements show the pace, and one result keeps us honest about what they mean.

    94.9%the AI alone< 34.5%people using it

    A test score is not a treatment

    In a 2026 study, AI models working alone named the right medical condition in 94.9% of test cases. People using those same models got there in fewer than 34.5%, no better than people without them. In that study the bottleneck was not capability. It was how the capability was put to work. That is why our agents run inside a process, with sources, a skeptic, a human reviewer and a lab, and why this page claims progress, not cures.

    Bean et al., Nature Medicine 32, 609–615 (2026)

    Open by default

    Everything in the open.

    Science moves faster when nobody has to take anyone's word for it. So the money, the work, the results and the decisions are all public, as they happen.

    Every dollar

    Follow the money, line by line.

    Every research fee lands in the treasury on-chain, and every dollar of compute is logged against the run that spent it. Audit both, any time: the treasury on Solscan.

    Cure-Fi treasurySolana
    Balance…Live
    • No research fees yetThe first coin's share lands here0 SOL

    Public log

    Watch the work as it happens.

    Every agent run

    Read the trail, not a summary.

    What each agent read, tried and concluded is logged and searchable. So are its mistakes.

    Every finding

    Published with its receipts.

    Results go out with the data and code behind them. Dead ends too: they save the next team months.

    Every decision

    Decided on-chain, in the open.

    Every coin names its disease and its research share at launch, locked on-chain. Coins with no disease feed the pool, which backs the diseases nobody has picked.

    Every tool

    Open data, down to every run.

    Every coin, fleet, run, verdict and dollar is published as plain JSON. Read it, check it, build on it: nobody has to take anyone's word for what the agents did.

    Fair questions

    Asked, and answered.

    Do the agents replace scientists?

    No. They take over the reading, the cross-checking and the number crunching that eat a researcher's week, so scientists can spend their time on judgement: which questions matter, which leads to pursue and what goes to the lab.

    Can AI agents really find a cure?

    Nobody can promise a cure, and we will not. What agents change is the odds: thousands of hypotheses explored instead of a handful, far more of the literature read than any team could manage, every hour of the day put to use. More shots on goal, better aimed.

    What can the agents not do?

    They cannot run an experiment at the bench, treat a patient or stand in for a clinical trial. They can also be confidently wrong. That is why every result has to get past a second agent whose only job is to break it, then a human reviewer, and, for the strongest leads, a real experiment in a lab.

    How do I know the work is real?

    You can check it. Every agent run, every result and every dollar spent is published. If a claim cannot be traced back to its data, it does not count.

    What happens when an idea fails?

    It gets published. A dead end that is written down saves every team after it from walking the same road. Most of science is ruling things out; we just refuse to hide that part.

    How is this different from a charity or a biotech?

    Most research charities fund grants and wait for papers. Most biotechs own their data and publish on their own schedule. Here every coin hires agents for one disease, the fleet grows with every coin, most of the legwork is done by agents, and every result is published, whoever it helps.

    Where does the money go?

    The launcher picks how much of a coin's creator fees goes to research, at least 40%, locked on-chain at launch. That share goes to the Cure-Fi treasury and pays for agents: every 0.2 SOL adds three to the coin's disease. Coins launched without a disease feed a shared pool for the diseases nobody has picked yet. Every movement is public.

    Are the numbers on this page real?

    Yes. Every figure in the evidence sections links to its source. Everything about Cure-Fi itself, the coins, fees, agents, runs and findings, is read live from Solana and from the public record at curefi.fun/data. Where there is nothing yet, the page says so rather than showing a guess.

    Community

    A cure is a team effort.

    Three groups, one obsession. Here is what each one brings to a project, and what each one gets back.

    Researchers

    You set the direction.

    The agents bring you ranked hypotheses with the evidence for and against. You decide which ones earn an experiment, and your name goes on every finding you guided.

    Developers

    You build on the record.

    Every run, verdict and cost is published as plain JSON, ready for your own dashboards, bots and analyses. Nothing to sign up for.

    Community

    You decide what gets cured next.

    Launch or back a coin for the disease that matters to you, and watch its fleet grow with every trade.

    Research fees by diseaseLive
    No coin yet0 SOL

    What each one gets back

    • Researchers

      A research team that never tires, and funding without the grant treadmill.

    • Developers

      An open record of real research to build on, from the very first run.

    • Community

      A straight line of sight from what they put in to the science it paid for.

    Get started

    Help fund the next cure.

    The next breakthrough is out there. Help us get to it first.

    Launch a project See how it works
    1. 1Pick the disease that matters to you
    2. 2Choose how much of its fees fund research
    3. 3Watch its fleet of agents grow
    Sources 35 references, each read at its source on 14 October 2026

    The thesis

    The workforce and the evidence

    Community funding and open science

    Open data

    The trajectory

    Figures about Cure-Fi itself are read live from Solana and the public record. Database counts change daily.