Insights

Latest analyses and perspectives on AI in chemistry, pharma, and biotech.

Shelf-life prediction with Arrhenius: from weeks of data to a shelf-life estimate in minutes

From a few weeks of accelerated stability data to a shelf-life estimate in minutes — on an amoxicillin example. Why the limiting parameter is not assay. And why the tool itself tells you when not to rely on extrapolation alone.

May 24, 2026

Use Case9 min read

Justify the worst-case marker of a shared facility in hours — not weeks

Cross-contamination risk assessment rethought: CovaSolv cleanability × CovaTox hazard surfaces the worst-case marker of a multi-product facility in hours instead of weeks. Why it is the combination, not a single axis — and why carbamazepine wins on the example roster, ahead of methotrexate and ethinylestradiol.

May 24, 2026

Use Case8 min read

ICH M7 automated: mutagenicity triage in seconds — a Losartan worked example

ICH M7 requires two (Q)SAR methods per impurity. On a Losartan worked example: 10 impurities classified in seconds, 5 probable mutagens identified — including the well-known azido impurity (AZBT). How CovaTox covers both methods in one call, and where the limits are.

May 24, 2026

Benchmark10 min read

Can AI predict toxicity? The CovaTox benchmark — measured honestly

CovaTox covers 42 ADMET / toxicity endpoints in one suite. In a true same-split, same-metric head-to-head against the TDC leaderboard, we sit a few points below the best dataset-specialized models on the AUROC endpoints — with one pipeline across everything. We name the weak spots openly.

May 24, 2026

Position11 min read

Agentic AI across the drug-development lifecycle, from discovery to scale-up

Agentic AI over MCP connects fragmented R&D data into one system. Example pipelines from discovery through CMC to manufacturing, and why the open tool layer decides the actual competitive edge.

May 24, 2026

Use Case7 min read

Designing carbamazepine crystallization, three steps with CovaSolv

Solvent screening, cooling curve, anti-solvent. Why acetone dissolves more at both temperatures but ethanol gives the higher yield, and how water as an anti-solvent crashes solubility by a factor of 220 without cooling a single degree.

May 23, 2026

Position9 min read

Context stuffing vs. tool calling: why many AI projects fail at the data architecture

More context does not make LLMs more reliable, the research shows the opposite. Three documented failure modes (lost in the middle, context rot, tool overload), why data architecture decides the outcome, and when tool calling is the more reliable choice.

May 22, 2026

MCP / Tech8 min read

RDKit over MCP: where the open-source toolkit stops, and where CovaSyn picks up

RDKit MCP servers give AI agents deterministic cheminformatics. Where the limit sits, no trained ML predictions, no hosting, no compliance, often too many generic tools, and what CovaSyn adds on top. With an honest recommendation for when a pure RDKit MCP server is enough.

May 22, 2026

Explainer6 min read

Why LLMs fail at chemistry, and what the tokenizer has to do with it

LLMs fail at counting rings. The reason is not missing chemistry knowledge but the tokenizer: SMILES strings shatter into disconnected fragments, and the molecular graph is lost. What the research shows, and how validated tools solve the problem.

May 22, 2026

Benchmark10 min read

CovaSolv benchmarked: predicting solubility at R² 0.92, and why the honest number is the more important one

CovaSolv predicts logS at R² 0.92 and RMSE 0.64, on 5,315 unseen molecular scaffolds. 78 % of predictions sit within 0.5 log units of the measured value. How close that is to the physical noise floor, honestly explained.

May 22, 2026

Benchmark8 min read

Gemini 3.5 Flash on chemistry: from 14 % to 76 %, why the cheapest model makes the biggest jump

Gemini 3.5 Flash reaches just 14 % on molecular reasoning without tools. Attach CovaSyn MCP and the same model jumps to 76 %, a 5.5× lift, the largest of all tested frontier models. Why the cheapest model benefits most, what it costs, and where the lift is not yet perfect.

May 22, 2026

Agent Setup11 min read

Improve Claude's Chemical AI Capabilities, Drug Discovery, Biologics, ICH M7 via MCP

Claude Haiku 4.5 reaches 21 percent on the ICLR 2026 chemistry benchmark. Attach the CovaSyn MCP server and it jumps to 85 percent. Same model, no fine-tuning, no prompt magic, just deterministic chemistry tools wired through the Model Context Protocol. This is the practical guide: how to wire it, what it unlocks for drug discovery and biologics, and the limits to know.

May 22, 2026

Agent Setup10 min read

AI Chemistry with Gemini, CovaSyn MCP Tool Calls for Drug Discovery and Biologics

Gemini's tool-calling reliability got pharma-usable in 2026. This guide shows how to attach CovaSyn's 130 deterministic chemistry tools to a Gemini agent for drug discovery, biologics, and process development, via the Gemini CLI and the Gemini API on Vertex AI. With the ICLR 2026 benchmark numbers and a worked tool call.

May 22, 2026

Agent Setup10 min read

Chemical MCP Tools in ChatGPT, Drug Discovery, Biologics and Regulatory Workflows

The ChatGPT App Directory and OpenAI's MCP support turned ChatGPT into a viable pharma-R&D agent in 2026. This is the practical guide for attaching CovaSyn's 130 chemical tools to a ChatGPT workflow, via the App Directory, Custom GPTs, the OpenAI API, and the new Apps SDK. With the ICLR 2026 benchmark numbers and a worked tool call.

May 22, 2026

Agent Setup11 min read

Chemistry MCP for Cursor, DeepSeek and Qwen, Open-Weights AI for Drug Discovery

Not every pharma team can run on Claude or GPT, IT security, data residency, or budget force open-weight stacks. This guide shows how to attach CovaSyn's 130 chemical MCP tools to Cursor, DeepSeek, Qwen and any LiteLLM-routed open-weight model. With concrete configs, the cost math, and where open-weight breaks down for chemistry.

May 22, 2026

Position8 min read

Data quality isn't the real bottleneck. Why 55 percent of biotech AI pilots actually fail.

In the Benchling Biotech AI Report 2026, 55 percent of 100 surveyed AI leaders name "poor data quality" as the top reason their pilots fail. The industry conclusion, "we need better data management", addresses the symptom, not the cause. A counter-thesis, three real failure modes, and a concrete proposal.

May 21, 2026

AI Scientist10 min read

The AI Scientist meets MCP: what the three Nature papers from May 19, 2026 mean for deterministic chemistry tools

FutureHouse (Robin), Google DeepMind (Co-Scientist) and DeepMind (ERA) published simultaneously in Nature on May 19, 2026: AI systems now generate hypotheses, design experiments, and optimize scientific software end-to-end. What the papers don't quite say: without a deterministic, validated computation layer the loop bottlenecks on human review. That's where the Model Context Protocol fits in. A reading frame.

May 20, 2026

Benchmark9 min read

14 % → 92 %: How CovaSyn scores on the ICLR 2026 chemistry benchmark

Klambauer's lab (JKU Linz) built a molecular-reasoning benchmark for ICLR 2026 that tests real chemistry instead of using LLM judges. Four frontier LLMs score 14 to 41 percent on it. With CovaSyn MCP attached, the same models reach 76 to 92 percent, three of them above 85. Four models, 12,540 responses, the full numbers including the gaps.

May 18, 2026

Market Overview12 min read

The 5 Leading Chemistry MCP Servers for Pharma R&D Compared (2026)

Aichemy, ChemMCP, CovaSyn, DIY Python stack, and OpenChem MCP, five ways to connect AI agents with chemistry, tox, and stability tools. A neutral overview of tool coverage, compliance, hosting, and pricing.

May 16, 2026

CovaSyn MCP

Run scientific tools right inside your AI agent.

Free tier: 100 credits per week. No credit card.

Try CovaSyn MCP →