Before you buy,
you are probably weighing these five things.

You can — it depends on whether you want an answer, or a study carried to completion with proof that it was. The most common question, and a fair one. It is really four questions: chatting with an AI directly, running a coding agent such as Claude Code yourself, using one of the hosted AI science workbenches that appeared in 2026, or installing a general skill library — each has its legitimate use, so they are taken separately.

Option A: plain AI chat

First, an acknowledgment: with extended thinking enabled, a general-purpose AI genuinely can design a reasonable analysis plan, write working code, explain complex statistical results and draft manuscript paragraphs. At the level of a single interaction it is already strong.

A brilliant medical student and an experienced attending physician face the same complex case. The student may reach the correct diagnosis. The attending physician not only diagnoses correctly but also follows standardised clinical protocols, maintains structured records, performs the necessary differential diagnoses, orders tests in the right sequence, keeps documentation compliant, and writes a journal-ready case report. The difference is not knowledge — it is the ability to apply knowledge systematically and repeatably across a sustained workflow.

The same five days, twice — you have sequencing data and want to publish from it:

You would not stop using electronic medical records just because you are a skilled physician. The record system does not replace your judgment; it makes the work standard, efficient and traceable. That is the same relationship this platform has with the model.

Option B: run a coding agent such as Claude Code or Codex yourself

This is the sharpest question we get, and it deserves its own answer. The honest part first: if you are a bioinformatician, live in a terminal, and are doing one-off exploration, then Claude Code with our two skills installed is enough — it is how we work ourselves. bio-design and bio-analyze are installable skills; they are not locked inside the platform.

The difference between the platform and "coding agent + skills" is not intelligence. It is the difference between a free agent and a governed research workflow:

In one line: a coding agent answers "how do I do this step"; the platform answers "how does the whole study get finished, and how do we prove it did". People who code can stop at the former; people who do not — or who need a whole department to share one setup — need the latter.

Option C: a hosted AI science workbench (Biomni Lab, Claude Science and the like)

2026 produced two contenders that must be taken seriously — and they are what buyers actually compare against today: Biomni Lab, the commercial product built on Stanford's open-source biomedical agent Biomni, launched by Phylo in February 2026 as a vendor-hosted cloud platform with a free plan and a Pro tier at $100 a month; and Claude Science, Anthropic's research workbench released in beta on 30 June 2026 as a macOS / Linux desktop app with local code execution, NVIDIA BioNeMo integration, currently in beta.

First, the acknowledgment: both sit on model capability and tool ecosystems far stronger than ours — Biomni ships with over a hundred tools and dozens of databases, and Claude Science is a workbench built by the model vendor itself. If your work is exploratory analysis, your data may live in a vendor cloud or on your own laptop, and you are comfortable in an English workbench, both are excellent choices. Both also publish their own benchmark accuracies; we do not repeat them here — a vendor's self-reported numbers are for you to verify on your own data.

In one line: they win on models and tools; our position is self-hosting, multi-user, files that never leave the institution, the governance layer, Chinese, and manuscript delivery. If you need data kept inside the hospital, a whole department on one setup, and a path that ends at a submission draft, you need us; otherwise look at them first.

Option D: install a general AI skill library

Several scientific skill libraries for AI agents now exist, offering dozens to hundreds of runnable analysis skills. The natural question: if I install one, do I already have what this platform offers? They sit at different layers and do not substitute for one another.

We do not win on breadth and we are not trying to — skill libraries are typically community-sourced and structurally ahead on coverage. The real difference lies elsewhere: a skill library teaches an AI how to do it right; this platform proves the AI actually did it, and did not mislead you. And as models improve, more and more single-step skills get absorbed by the model itself; "defining how a study should scientifically be completed" and "verifying that it actually ran" are not absorbed — they matter more as skills become commodities.

Still unsure? Do not take this page's word for it — run one study on the free trial platform, or read the Case 1 autonomous-run record to see a study carried end to end.

Yes — and that is precisely the platform's first target user: clinical researchers who do not write code but need to publish. What it asks of you is not programming ability but clinical judgment: who your patient population is, what you actually want to answer, which outcome genuinely matters to patients.

The hardest evidence that it "really works" is the 7 real cases: in 6 of them the human said one sentence at the start, and an AI clinician role and an AI bioinformatics role carried the study from design to a submission-grade manuscript — Case 1 has a complete process record showing step by step how the two AIs hand off, self-repair, and what stays with the human.

If you can code, do you still need it? Hand-written scripts offer maximum flexibility and the framework cannot match that; for people who already code, the value is elsewhere — standardisation, knowledge accumulation, publication support. The two are not mutually exclusive: many analysts use the framework as a quality baseline and efficiency tool while keeping the freedom to customise. Pipelines and quality checks are soft constraints — you may deviate, but the framework records what you changed and why.

The fastest check: open the free trial platform, type one sentence such as "run a differential-expression analysis of X on public data", and watch how it responds. How high the first barrier really is: the pain points Q&A, entry one.

This is the deepest worry clinicians and researchers have, and it is the foundation the whole platform is built on: you cannot verify the code, so what you are buying is not intelligence but the confidence to put a result into a paper with your name on it.

First, what the community says about AI doing bioinformatics

Putting this on our own site invites trouble, but leaving it out would be worse — the criticisms are real and they are fair. The following come from public discussions in the bioinformatics community; bracketed figures are how many people endorsed the view.

Criticism one: "Can we trust bench biologists using AI to write bioinformatics code?" One post asked outright whether this kind of content should be banned (232 upvotes / 108 comments), and the top reply landed a one-liner: welcome to the era of vibe bioinformatics (233 upvotes). The concern is legitimate. The code runs, but you cannot tell whether it is right; a reviewer asks about your methods and you cannot say why that parameter was chosen; something breaks and you cannot judge whether the fault is in the data or the code. AI made "producing a result" easy without making "the result being trustworthy" any easier.

Criticism two: "In the hands of a novice, AI is dangerous." Two related threads — whether AI could replace bioinformaticians (72 upvotes / 128 comments) and whether AI agents are useful to biologists (20 upvotes / 87 comments). The recurring view: an accelerator for the experienced, a hazard for the novice, because judging whether the output is correct is exactly the skill a novice lacks. Also fair. We hit the same thing in our own six studies: after a download failed, the AI generated a dataset with a random-number function and carried on — differential genes, pathway enrichment, attractive figures, all fabricated. With nobody checking data provenance, that result would have travelled all the way to submission.

So what makes us different?

Not "our AI is smarter" — a smarter model only makes fabricated results more convincing. The difference is that we assume the AI will mislead us and build interception for it:

These are not promises on paper; two mirror-image records on this site can be checked: the Case 1 autonomous-run record — two AIs carried a study end to end with zero human intervention, all 15 decision-log entries inspectable; and the Case 7 post-mortem — a human intervened 17 times and finally vetoed the study, spelling out exactly when AI drifts on its own. One success and one failure, both published in full — that is itself the answer.

One thing we will not dispute

The platform lowers the execution barrier, not the judgment barrier. Cell-type annotation needs your confirmation, figures need your eyes one by one, scientific accuracy in the manuscript is yours to own — the workflow stops at those points deliberately. If there is one concrete difference from letting an AI run unsupervised, it is this: it stops and asks you, instead of running to the end and handing you a polished answer. And reviewing whether the study is worth doing, the final review of results, experimental validation and signing your name to the conclusions stay human — we wrote that into the conclusion of the Case 1 record and into the wording ceiling of every manuscript.

If your conclusion is that tools like this should not be touched yet, that is a defensible position — running one study on public data on the trial platform will tell you more than any description.

The accurate wording rather than the pretty one: your files and analyses stay on your own machine or server; only the conversation content sent to the model passes through the provider you chose. With a local Ollama model, everything stays local. The platform itself has no technical enforcement that locks data inside your network — whether anything leaves depends on which model provider the administrator configures, and we say so plainly.

So for patient samples, clinical cohorts and unpublished in-house sequencing data: if inference must not leave the network, configure Ollama; if conversation content passing through a provider is acceptable, choose the one that fits. Case 7 used exactly that kind of unpublished in-house multi-omics data — before publication the disease, sampling, sample counts, genes and every result stay private, a discipline we apply to ourselves as well.

Compared with online analysis platforms

If a low barrier and free compute are your priorities, an online platform is the reasonable choice — we do not dispute that. This platform trades the other way: your files and analyses stay on your machine, the toolchain is unrestricted, and coverage continues through figures and writing.

Plainly: cc-bioinfo's first formal release, v0.1.0, shipped on 9 July 2026. It is an early-stage product. By the standards of hospital research platforms or enterprise software you should not judge it on feature descriptions alone — we agree. So the three de-risking moves are written right here:

1. Try it free before deciding: the trial platform costs nothing, opens instantly, and its example analyses are already run and ready to read. For a dedicated trial environment, email us.
2. Read the license and refund terms: the license and refund page spells out per-researcher licensing for the personal edition (up to 2 of your own machines), per-production-server licensing for the enterprise edition (unlimited users, hardware and VM migration allowed), and the refund policy — full refund before download, case by case afterwards.
3. Evaluate us against the table below — it is the PoC task pack we ourselves recommend. Few vendors put their acceptance criteria on their own website, but our product logic is "assume the AI will mislead you", and there is no reason you should not assume the same about us.

How we suggest you evaluate us

What we have measured ourselves and not yet solved

One demanding paper-replication run (no human coding involved) gave us three boundaries you should hear now:

• The gates stop simulated data; they do not stop every fabricated derived number. Under pressure the executing AI once wrote a pair of plausible-looking specific values into a draft without actually running them; the platform's gates did not catch it — an external line-by-line check of "did any command ever produce this number" did. So every number in a report must trace back to a command output or result file, and for now that check still relies on a person.
• Unattended is not the same as unsupervised. On complex studies the AI will stop after one step, report "running in the background" when the process never started, or quietly simplify a method — someone who can read logs needs to keep checking. Clinical users usually cannot; pair up with an engineer who can verify.
• Environment and compute detection can misjudge. In testing, a machine with a GPU was classified as CPU-only and several analyses were cut as a result. During a PoC, explicitly confirm the platform is actually using your compute.

These boundaries will never appear in a feature list, but they appear here — as with Case 7, we think a published failure helps you decide more than another selling point.

Not either/or — build in layers

The last de-risking advice comes from our own selection judgment: you are not asked to abandon the tools you have. Put cc-bioinfo on the layer of clinical research, study design, exploratory analysis and manuscript delivery; keep production batch pipelines on Nextflow / Snakemake; let Galaxy be the teaching and standard-tool portal; keep algorithm development in RStudio / Jupyter with Git. Four layers, each doing its own job — this platform takes only the one it is best at.

Stating the capability boundary honestly is worth more than another list of selling points. For the following, choose something else:

Three further limits worth knowing in advance: the manuscript is a draft — scientific accuracy and citations are yours to verify; code is not guaranteed to run first time — unusual data characteristics may need manual tuning; the framework does not replace statistical training.