GlowSky

For medicinal chemists and drug-discovery researchers. Describe a design idea in plain language, and a chemistry-aware assistant plans the work and runs real, checked tools to do it. The AI reasons and explains; the chemistry (RDKit, docking, ADMET) does the actual calculating. Every molecule is a real, saved, viewable object — never a made-up text string. And you bring your own AI model.

Early access. GlowSky is built and works end-to-end — even fully offline against a built-in test model — and you can run it on your own machines. A fully managed, hosted version arrives with our first design-partner lab. It's the platform's AI sibling: a separate workspace that shares your Nakitte sign-in but keeps its own chemistry-focused way of working.

How a design run works

• Plan, generate, profile, filter, rank and explain — in one run — the AI plans and explains in words, while every chemistry step runs through real, checked tools, with a full record of what happened.
• Start from a seed molecule to generate close analogues, then check them for drug-likeness against the goals you set for the run.
• Keep the conversation going — it remembers the molecule you're working on across messages, so "now make them more polar" just works.
• Watch results stream in live — the plan, each candidate, the ranking and the explanation, as they happen.

Real chemistry — never a made-up molecule

• Every structure is checked before it's trusted — whether you typed it, imported it, or the AI proposed it, it's parsed and cleaned up before it's ever shown, saved or used.
• A full set of real RDKit tools — descriptors, drug-likeness, PAINS/BRENK alerts, fingerprints and similarity, substructure search, scaffolds, synthesizability and 3D conformers.

Medicinal-chemistry know-how

The expert layer that makes GlowSky more than a property calculator — and it's all real calculation, with no AI in the maths.

• Spot what a small change does (matched molecular pairs) — it finds molecules that differ by one group and tells you the average effect ("H→F lowers logP by 0.4 on average, n=6"), ordered by how much evidence there is.
• Score drug-likeness against the profile you want — for oral, lead-like or fragment goals, and run the well-known rules (Lipinski, Veber, Ghose, Egan, Muegge, lead-likeness, Rule of 3), each one a clear pass or fail with the reasons.
• A first take on how to make it — common reaction disconnections (amide coupling, Suzuki, reductive amination, and more), honestly presented as a guide rather than a full route planner.
• Swap groups and hop scaffolds — well-known, curated replacements (acid → tetrazole, ester → amide, and others), each result checked before it's shown.

A chemistry toolkit you can also use on its own

Each of these is a tool the assistant can call — and one you can run by hand too.

• Similarity scoring (one pair, or ranking a whole set) using standard molecular fingerprints.
• Substructure search and scaffold extraction.
• A synthesizability score and real 3D conformer generation (with energies) that drives the 3D viewer.
• Import and export SMILES, CSV and SDF — a partly-bad file still loads the good rows, and you can compare two molecules side by side.

ADMET and docking — real, and honest about confidence

• Built-in offline ADMET estimates — solubility, blood-brain-barrier, logD, hERG, CYP3A4, metabolic stability and protein binding — each one labelled with how it was worked out and how confident it is.
• Plug in your own ADMET model or a docking engine, and it becomes available to the assistant — and out of the box it would rather say "not set up" than make a number up.
• AutoDock Vina docking (when you turn it on) — real 3D poses, with the target file kept to a folder you allow; until then it simply stays off.

One engine behind it all

• One place runs every tool, picks the right version, sends quick jobs and slow jobs the right way, checks every result and keeps a record.
• Work is computed once and remembered — and kept private to your own organisation, since your molecules are your IP.
• Tools are versioned and can be pinned so a project gives the same answer every time.
• Run a tool over many molecules at once and watch progress, with any failures reported rather than quietly dropped.

Bring your own AI model

• Use whichever model you trust, with a built-in offline option — so you control which model sees your work and where any spend goes.
• Connect Anthropic, OpenAI, Groq or a local model — your key is used only when needed and never logged — or run the whole thing offline with no key at all.
• Send different jobs to different models, with your own preferences on top of the defaults.
• Your provider keys are encrypted — only a masked hint is ever shown back.

Open and extensible

• Add your own model as a sandboxed tool — package it as a container, and GlowSky runs it safely (no network, read-only, locked down) right alongside the built-in tools.
• Register a custom tool with a small manifest — no GlowSky code change, and it gets the same checks, caching and records as everything else.
• The same approach works beyond chemistry: the Nakitte ULD line's AI tools plug in the same way — for example cargo dimensioning sitting next to the chemistry tools.

One sign-in, kept separate per team

• Sign in with your Nakitte account — the same login as the rest of the platform. One sign-in across the family.
• Organisations, projects and molecule libraries — with roles, your saved collections, your encrypted model keys and a full history — all kept private to your team.

Reproducible and easy to share

• Export a run as a Jupyter notebook — self-contained code that recreates the numbers and re-applies the filters.
• Or a readable written report — the goal, the plan, a table of candidates and the assistant's explanation.

On the desktop

• A desktop app — candidates stream in as they're checked, with 2D and 3D views, library import/export, a docking-pose viewer and your model settings.
• A library workspace for your projects and molecule collections — import SMILES / CSV / SDF, browse, and export.
• A docking screen with a 3D pose viewer, shown on a real HIV-protease example.
• A molecule inspector — drug-likeness scores, a first synthesis route and PAINS/BRENK alerts at a glance.
• A tool explorer that turns any tool into a simple form, a quick command palette, and a built-in structure editor for drawing molecules, all working offline.
• Settings — sign in, manage your encrypted model keys, and choose which model handles which job.

Built to be trusted

• A broad set of tests runs entirely offline and checks the chemistry and the design run end to end — from matched-molecular-pair analysis and the drug-likeness checks to the live design stream and platform sign-in.

Coming soon

• A fully managed, hosted versionComing soon — GlowSky runs on your own machines today; the hosted option arrives with our first design-partner lab.
• Email / single sign-on loginComing soon — the sign-in is in place; a friendly login screen arrives with the hosted version.

See also

• All products — the rest of the Nakitte Carbon family.