Talk to your data.
Environment, code and execution handled. The scientific judgment stays yours.

Try it free first, buy it only if it fits. Three channels for three situations:

Prices are in USD, sold through Gumroad. Individual users in mainland China: see the Chinese buying guide.

Personal vs Enterprise at a glance

The platform's product introduction and public documentation are available in the public cc-bioinfo repository (README, changelog, feature notes); deployment packages and installation support are provided after purchase. Please read the license and refund terms before buying.

cc-bioinfo is not a chat wrapper and not yet another skill: it is a multi-user analysis workbench installed on your institution's server or your own computer, on which the two core skills (bio-design, bio-analyze) run. The skills can perform on other stages; the difference is the stage itself. Below, one row per thing you may be using today — each starts with where the other side is the better choice.

It is a server-side, multi-user platform: researchers use it through a browser with nothing installed locally; closing the browser interrupts nothing — running analyses, scheduled jobs and IM delivery all continue, so a multi-hour end-to-end analysis can finish unattended. Users are isolated from each other at the Linux UID level with PAM login, sharing one pool of compute and model access without seeing each other's work.

The item-by-item comparison — including skill libraries, hand-written scripts and the cases where this platform is the wrong choice — is on How to choose.

Features in the order a study uses them rather than as a list: ① frame → ② build the environment → ③ execute → ④ monitor (including reviewing every code change the AI makes) → ⑤ extend. The deliverables themselves — figures, manuscript draft, submission package — come out of bio-analyze in step ③.

• ① Frame · bio-design study designThe upstream design skill: literature decomposition plus heuristic design dialogue, backed by 382 decomposed cases and 6,235 structured decision patterns.
• ② Environment · Zero environment setupBuilds a per-project conda environment automatically, installs missing packages on demand, and bootstraps miniconda from scratch when conda is absent — you never have to survive dependency hell.
• ② Environment · Multi-model ecosystemNatively optimized for the Claude API; unified routing to AWS Bedrock, DeepSeek, Kimi and Ollama — with Ollama, inference runs entirely on your own machine.
• ② Environment · Three-tier configurationProject tier over user-global tier over framework defaults. Lab-specific marker gene lists merge with built-in knowledge instead of replacing it.
• ③ Execute · bio-analyze autonomous executionThe downstream execution skill: a Step 0-7 workflow, multi-hour analyses completed unattended, all the way to an SCI manuscript draft.
• ③ Execute · Chat-to-code workflowDescribe the task in natural language; the system writes code, installs missing dependencies and provisions the runtime. Hand it a stack trace and it diagnoses the cause and returns corrected code.
• ③ Execute · Built-in browser terminalA per-user isolated tmux terminal for shell commands, package installation and environment management — no SSH client required.
• ④ Monitor · Multi-channel IM integrationWeChat, DingTalk, Telegram and Feishu adapters all pass end-to-end tests: permission approval, project switching and streaming result cards let you decide from a phone between clinics.
• ④ Monitor · Immersive IDE experienceA full browser-native Web UI: AI conversation on the left, code changes and diffs on the right; multi-project organization with session history and branching.
• ⑤ Extend · Plugin ecosystemInstall community plugins from the Web UI or CLI: PubMed search, scRNA-seq QC, Open Targets, Consensus and more.

cc-bioinfo uses a three-layer multi-user architecture: an nginx reverse proxy with Lua, a Go authentication service backed by Linux PAM with sessions and audit logging, and one isolated Bun process per user managed by systemd template units.

• Process-level isolation: each user session runs under its own Linux UID with PrivateTmp, memory limits and CPU quotas.
• Unix socket routing: requests route through a per-user socket, so users cannot reach each other's sessions — verified by a real cross-user isolation test with 2 users signed in concurrently.
• Institutional account authentication: the authentication service speaks Linux PAM directly, so researchers log in with existing institutional credentials and no separate account system is needed.
• Server-resident persistence: user processes are systemd-managed and independent of browser connections; services recover automatically after a host reboot.

On data residency, stated honestly: raw data is only read by the local R/Python environment on the server. Whether any data leaves the intranet depends on the LLM provider your administrator configures. A self-hosted LLM such as Ollama can be configured for full data residency — but the platform itself does not technically enforce it, and we say so plainly rather than implying a guarantee it cannot make.

1. Step 1 · Open the trial platform: test.cc-bioinfo.com is free and needs no sign-in; for a dedicated trial environment email tangmoogmoogtang@gmail.com. You talk to it in the browser. Two good first moves: drop in a reference paper from your field for bio-design to decompose, or take a public GEO dataset and let bio-analyze run QC and clustering.
2. Step 2 · Pick an edition: personal edition for individual use, enterprise edition for a department or institution. See the purchase section.
3. Step 3 · Deploy and activate: for the personal edition, follow the setup guide — three videos cover the whole process. Enterprise edition deploys on your own servers — the setup wizard installs every dependency, configures TLS and runs 17 self-checks. Deployment packages and installation support are provided after purchase.

The two columns differ widely because they measure different things. The enterprise figures cover the platform itself; the actual analysis runs on institutional compute. On the personal edition everything happens on your machine — platform, conda environments, raw data and computation — so size it for real single-cell and spatial work. The setup guide covers obtaining the image and installing it.

The six studies below were completed end to end on cc-bioinfo by an AI clinician role and an AI bioinformatics role working together; the human said one sentence at the start. All use public real data, including honestly reported negative results. Full process archives are on the case studies page.

• Mendelian randomization
  statistical genetics
  Causal architecture of AF → HF and its protein mediatorsOn IEU OpenGWAS summary statistics: AF→HF OR 1.34 (p=3×10⁻⁵⁹), all 1,915 plasma proteins screened, zero mediators — one sentence in, an 8,060-word English manuscript out.
• Single-cell + spatial
  cardiovascular
  A five-layer evidence loop on carotid-plaque SMC stability20,909 SMC-lineage cells plus 120,164 cells of Xenium spatial validation; when the data did not support the premise, the AI rebuilt the paper's thesis from scratch.
• Pharmacogenomics
  comparative study
  CAVD vs CAD causal maps and drug repurposing300 targets, two outcomes, MR across seven parallel evidence lines in 26 hours — a genetic explanation of why statins fail against valve calcification.
• Mechanistic pharmacology
  molecular docking
  Sevoflurane × myocardial ischaemia-reperfusion × ferroptosisHit a preregistered gate and turned honestly: every docking score fell short, so the conclusion was framed as "not single-target inhibition".
• Evidence-based medicine
  umbrella review
  A health check of the HFpEF drug-therapy evidence base102 meta-analyses: 96.1% rated Critically Low on AMSTAR-2, 16.7% corpus overlap — "the field needs better meta-analyses, not more".
• Reproductive medicine
  self-chosen topic
  Three successive rounds on the endometrial window of implantation in recurrent implantation failureThe AI set its own question, narrowed it over three rounds and closed the door honestly three times — each round ruled out a widely assumed mechanism.

Two mirror-image process records: the Case 1 autonomous-run record — one human sentence, two AIs, zero intervention, an 8,060-word submission draft, and how the four integrity mechanisms work and what stays human; and the Case 7 post-mortem — 17 human interventions and a final veto, showing when AI drifts on its own. One success and one failure, both published in full.