How it works
An honest account of what the box on the front page does today, what it computes from your own input, where it shows you a worked example instead, and how to get the same tools inside the editor you already work in.
What happens when you hit run
-
It matches keywords. There is no language model here.
Your line is scanned for words like
primer,Golden Gate,Opentrons,GFPorbarcode, and for anything that looks like pasted DNA. First match wins. That is the whole of the "understanding" — it is a switchboard, not a brain, and it will say so rather than guess if nothing matches. -
It calls a real function.
Whichever tool matched is the same code that runs for everyone, installed on your machine by
apt install bioand exposed to your agent over MCP. Nothing is special-cased for the website. -
It shows you what came back, unedited.
Every number on the answer page was computed. No part of it is written prose generated to sound plausible. If a melting temperature says 58.4 °C, something calculated 58.4 °C.
What is computed from your input, and what is a sample
This is the distinction worth knowing before you judge the output. Most tools work on whatever you paste. The protocol generator works on your parts if you name them, and shows a worked example if you do not — and it says which of the two you got.
| Ask for | Uses your input | What it actually does |
|---|---|---|
| Primers "design a primer for this" |
Yes | Designs a real pair against your template. Nearest-neighbour melting temperatures from SantaLucia's 1998 unified parameters at 500 nM primer and 50 mM Na+, with GC content, the melting-temperature difference between the two, product length, and warnings for runs, clamps and self-complementarity. |
| A pasted sequence paste DNA, say nothing |
Yes | Length, GC content, longest homopolymer run, real BsaI site positions, a primer pair if it is long enough, and a scan for any sudo.bio barcode embedded in it. |
| Golden Gate check "check my overhangs" |
Yes | Finds the actual Type IIS sites in your sequence and tells you which need domesticating. The proposed fusion-site set is screened so none is palindromic, none is the reverse complement of another, and all are at least two bases apart — the single-mismatch mis-ligation that Potapov et al. 2018 showed dominates failed assemblies. |
| Provenance trace paste a sequenced read |
Yes | Really decodes the barcode, verifies its checksum and resolves it to the commit that produced the material. |
| A brighter protein "I need a brighter GFP" |
Partly | Reads the protein you named and looks it up in a curated table of published brightness, excitation and emission maxima, oligomeric state and photostability. The table is fixed reference data, not a prediction — which is the point. It will usually tell you not to run a screen. |
| Robot protocol "write me Opentrons code" |
If you name your parts | Name them in a list — golden gate with pTet, B0034, sfGFP and B0015 for 24
reactions — and the deck, tube positions, master mix and reaction count are built
around them. Name none and you get four placeholders, clearly labelled as such. Either way the
validation is real: every volume against the fitted pipette's range, no deck slot booked
twice, enough tips loaded, trash left clear. |
| Plate screen "how do I run a screen" |
Not yet | Standard deck layout and the controls that decide whether you get data or noise. This one is guidance, the same for everyone. Paste a plate reader export and the analysis is real — every well scored by median and absolute deviation. |
Why the failures are kept
When a result is committed, every well is recorded — the duds alongside the hits. That is the difference between "nobody has tried this" and "somebody tried this and it did not work", and only one of those is worth knowing before you spend a month on it.
The store is an ordinary bare git repository on disk. It is append-only by construction:
there is no code path that edits or deletes a committed record, because no such function
exists to call. git log works on it directly, and you can walk away with
everything.
The barcode
This is the part that has no equivalent elsewhere. When you commit a construct you get a 32-base DNA barcode that encodes the commit identifier, with a fixed anchor at each end and a checksum in the middle. You synthesise it into the plasmid.
Two years later someone finds a tube in the back of a −80 with a label nobody can read. They sequence it. The barcode comes back out of the read — tolerant of the mess real sequencing produces — and resolves to the commit that made the material: what it was, who made it, what it was built from, and every result ever attached to it.
The sample carries its own provenance. Not a label that refers to a database that may still exist, but the record itself, in the DNA.
Try it: paste a read into the provenance trace.
Using it where you actually work
The box deliberately has no language understanding. That belongs in the agent you already use, which can reach every one of these tools directly. This is the way the tools were meant to be used — the box is just a way to see them without installing anything first.
-
Install the command
One line, and the full instructions are on the install page.
-
Point your agent at it
Use the Install MCP button at the top of any page. Cursor is one click; Claude Code and everything else get the exact line or config block they accept.
-
Ask in your own words, in a conversation you were already having
The agent supplies the language understanding; these tools supply the arithmetic and the record. It can design, plan, emit the robot code and commit the result without you leaving the editor.
Things worth asking your agent once it is connected:
- "Design primers for this plasmid, then commit the design to a new repository."
- "I'm assembling pTet, B0034, sfGFP and B0015 — check the overhangs and write me the Opentrons protocol for 24 reactions."
- "Here's the plate reader export. Score it and commit every well, then tell me which clones beat the parent."
- "What is this sequence? Trace it back."
What it does not do, so you are not surprised
- It does not predict which mutation will work. For brightness it will usually tell you not to run the screen and to swap in a published protein instead, which is cheaper and almost always better.
- Tecan and Hamilton code is emitted at firmware level through PyLabRobot, which the manufacturers do not endorse and which may affect your warranty. Protocols are designed on an Opentrons OT-2 deck, so moving them needs labware substitutions — each one is listed with the script rather than hidden. For hardware PyLabRobot has no backend for, such as a Tecan Fluent, you get the transferable plan and an honest refusal to fake a script.
- Validated is not verified. Run generated protocols through
opentrons_simulatebefore the deck is loaded. - The box understands keywords, not sentences. Phrase something unusually and it will tell you it did not understand rather than invent an answer. Connect an agent and that limit goes away.
- There is no cloud and no lock-in. Everything runs on your machine if you want it to.
Back to the ask box, browse the repositories, or read the install instructions.