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

Lead

ChatGPT is the most-reached LLM in pharma R&D simply because everyone already has a ChatGPT Plus account. In 2026 it also became the most agent-capable consumer surface, thanks to the App Directory and OpenAI's MCP support. The CovaSyn chemical MCP tools attach to ChatGPT three different ways depending on what you want: zero-setup discovery via the App Directory, custom flows via a Custom GPT, and full programmatic control via the OpenAI API. Here's how to set up each, what each one is good at, and the limits.

What are chemical MCP tools?

Chemical MCP tools are deterministic cheminformatics functions — ADMET, druglikeness, ICH M7 mutagenicity triage, NMR and mass-spec interpretation, ICH Q1 stability, retrosynthesis hints, biologics developability — exposed to an LLM agent through the Model Context Protocol. The LLM calls them, gets back validated, version-pinned, audit-logged results, and reasons over them. The Model Context Protocol itself is the agent-tool standard introduced by Anthropic in 2024 and jointly maintained by Anthropic, OpenAI, and the open-source ecosystem since 2025.

For ChatGPT specifically, MCP support landed cleanly in the App Directory model: an MCP server published as a ChatGPT App appears as a connector in the user's ChatGPT settings; ChatGPT routes tool calls through it.

Why ChatGPT matters for chemistry agents

Three reasons to wire CovaSyn into ChatGPT, regardless of whether Claude is your primary stack.

First, distribution. The user already has a ChatGPT subscription. No new login, no new tool to learn. For med-chem teams that haven't yet adopted Claude, ChatGPT is the path of least resistance to a working agent stack.

Second, the App Directory model. A CovaSyn listing in the ChatGPT App Directory makes the chemistry layer one search-and-click away for any ChatGPT user. The discovery loop is shorter than for Claude Desktop or Cursor.

Third, multi-modal reasoning over chemistry assets. GPT-5.5 reads a hand-drawn molecular structure, a mass-spec spectrum image, or a TLC plate photo, and combines it with a CovaSyn tool call. That image-plus-structure path is the strongest ChatGPT-specific capability for pharma R&D.

On the ICLR 2026 MolecularIQ benchmark, GPT-5.5 reaches 22.29 percent baseline accuracy on chemistry tasks and 89.92 percent with CovaSyn MCP attached — a 4.03× lift, structurally identical to what Claude Haiku does. Same lever, same architecture, different ergonomics. Full methodology.

How CovaSyn connects chemistry to ChatGPT

CovaSyn exposes 130 deterministic chemistry tools as an MCP server. ChatGPT sees them in the same shape it sees any other connector: a list of tools with names, descriptions and JSON schemas. The agent picks the right tool, calls it, parses the structured response, and replies.

Coverage by domain:

• Drug discovery — ADMET, druglikeness, scaffold analysis, BRICS, similarity, ICH M7 triage, off-target screening
• Biologics — antibody developability, peptide profiling, ADC linker design, mRNA / oligo design, immunogenicity, viscosity
• Process development — design of experiments, response surface, solubility, crystallization, formulation
• Analytical — mass spec identification and fragmentation, NMR 1D/2D, retention-time prediction

A worked tool call

Concrete: a synthetic chemist asks ChatGPT about ADMET for a drug-like compound before committing to a synthesis route.

{
  "tool": "covachem_adme",
  "arguments": {
    "smiles": "Cc1ccc(NC(=O)c2ccc(F)cc2)cc1S(=O)(=O)N",
    "endpoints": ["caco2", "ppb", "logd74", "cyp3a4", "herg"]
  }
}

Response from the tool:

{
  "results": {
    "caco2_logpapp": -4.92,
    "ppb_percent": 88.3,
    "logd_7_4": 1.27,
    "cyp3a4_inhibition_prob": 0.42,
    "herg_inhibition_prob": 0.18
  },
  "confidence": "high",
  "engine_versions": { "covachem": "v3.1.0" },
  "audit_id": "ca-2026-05-22-7b6a5c4d",
  "timestamp": "2026-05-22T11:14:08Z"
}

GPT-5.5 takes the structured payload, summarises permeability as "moderate Caco-2", ppb as "high protein binding, expect lower free fraction", and flags moderate CYP3A4-inhibition risk. The user gets the chemistry interpretation; the audit log keeps the numbers.

Cost: cents per compound, accuracy at scale

For a typical ADMET screen on a 1,000-compound triage batch:

| Path | Cost per compound | Accuracy | Audit-fit | |---|---|---|---| | Manual ADMET review by chemist | €15–50 | high but slow | yes, by hand | | GPT-5.5 alone (no MCP) | €0.03 | ~22% on hard tasks | no, not reproducible | | GPT-5.5 + CovaSyn MCP | €0.10 | ~90% | yes, audit-logged | | Outsourced CRO ADMET panel | €80–300 | high | yes |

The €0.10 row is the one that changes the workflow. A pharma team can run an ADMET screen 1,000× per week, not once per quarter.

Where this breaks down

• Tool-call latency in ChatGPT App. Each tool call adds ~1–3 seconds of round-trip latency on the consumer ChatGPT surface. For multi-step workflows (8+ tool calls), this stacks. For batch screening, run via the API instead.
• App Directory review timing. When OpenAI changes App Directory review processes, listings can take longer to go live. The connector-via-URL path keeps you unblocked while the listing is pending.
• Regulated workflows on consumer ChatGPT. Consumer ChatGPT runs on OpenAI's US infrastructure. For EU-Annex-11 contexts route through the API with appropriate region choices, or use the self-hosted CovaSyn container behind your own LLM gateway.

Getting started — ChatGPT App Directory

The zero-friction path for end users.

1. Open ChatGPT → Settings → Connectors 2. Search "CovaSyn" 3. Click "Add" and authorise with your CovaSyn account (OAuth) 4. Tool list appears immediately; start a new chat to use them

Free tier (100 credits/week) is enough to evaluate.

Getting started — Custom GPT

For team-specific workflows.

```text Create a new Custom GPT in ChatGPT → Actions tab → Add Action → Authentication: API Key (CovaSyn key) → URL: https://workspace.covasyn.com/mcp/openapi.json → ChatGPT auto-imports the tool list

Instructions: "You are a med-chem co-pilot. For every compound the user mentions, run druglikeness and ADMET via the CovaSyn tools. Cite the audit_id in your response so QA can trace it." ```

Publish privately to the team workspace. Done.

Getting started — OpenAI API

For your own agent stack.

```python from openai import OpenAI from mcp.client.stdio import stdio_client

client = OpenAI() async with stdio_client("npx", "-y", "@covasyn/mcp-client", env={"COVASYN_API_KEY": "sk-cova-…"}) as mcp: tools = [t.to_openai_function() for t in await mcp.list_tools()] response = client.chat.completions.create( model="gpt-5.5", tools=tools, messages=[{"role": "user", "content": "Run ICH M7 on compound X"}], ) ```

For Self-Hosted-Container setups, point the MCP client at your internal CovaSyn instance. Same protocol.

FAQ

Sources

• ICLR 2026 MolecularIQ benchmark: covasyn.com/en/benchmark.
• AI for chemistry pillar: covasyn.com/en/ai-for-chemistry.
• Comparison of chemistry MCP servers: The 5 leading chemistry MCP servers.
• Bottleneck position piece: Data quality isn't the real bottleneck.