How to use

You don't call tools. You ask.

Once SciAgentKit is wired into your agent, you describe the goal in plain language. The agent picks the right MCP tools, runs them in the correct order, and hands back real files — poses, scores, trajectories, and a report. Here is exactly what to say and what comes back.

First, wire it in (once)

Paste this to your agent (Claude, Cursor, Gemini, or Codex). It installs SciAgentKit and registers the MCP server for you — full details on the Get started section.

prompt → your AI agent
Install SciAgentKit and set it up as an MCP server for this project.
Run: pip install "sciagentkit[mcp]" then `sciagent init-agent`.
Then restart so you can call its scientific tools.

Example prompts & what happens

Each prompt below is real natural language. The agent resolves it to a chain of SciAgentKit tools — you don't name them yourself.

1

Triage a molecule library

You“Profile my_library.smi and flag anything with toxicity or PAINS liabilities.”
Agent calls
analyze_library→adme_tox
You get
  • Cleaned, deduplicated library with property distributions
  • adme_tox.csv — ESOL solubility, drug-likeness rules, GI/BBB flags, PAINS/Brenk alerts
  • A summary of which molecules to drop and why
2

Dock ligands to a named target

You“Dock the ligands in ligands.smi to EGFR and give me the top hits.”
Agent calls
protein_select→ fetch_pdb→ prepare_ligands→ detect_pocket→ prepare_receptor_docking→ run_docking→ select_top_hits
You get
  • The chosen PDB structure and its ranking rationale
  • Docked poses (.pdbqt) and docking_scores.csv
  • top_hits.csv — target-aware ranking (docking + drug-likeness + toxicity)

Tip: “dock … to EGFR” one-shots the same chain via the dock_to_target tool, which stops gracefully if Open Babel or a docking engine isn't installed.

3

Set up MD and analyze the trajectory

You“Run molecular dynamics for the best pose and analyze the trajectory.”
Agent calls
assemble_docked_complex→ ligand_md_plan→ ligand_md_run→ analyze_trajectory
You get
  • A fully-specified OpenMM/OpenFF job (solvation + NVT/NPT/production) — the agent runs it with ligand_md_run
  • Backbone RMSD, per-residue RMSF, and a ProLIF interaction fingerprint from the resulting trajectory
  • Figures for each metric, computed by MDAnalysis/ProLIF (never estimated)
4

Run the whole pipeline & get a report

You“Run the full screening project for BACE1 with ligands.smi, then write me a report with wet-lab recommendations.”
Agent calls
target_screen_project→ select_top_hits→ wet_lab_plan→ write_report
You get
  • A structured report.pdf (11 sections: summary → docking → MD → wet-lab → limitations)
  • A staged wet-lab plan tailored to the target class (BACE1 → CNS/BBB stage included)
  • run_manifest.json — the signed, reproducible record of the whole run

Prompt cookbook

A quick reference — say the goal, the agent handles the tools.

Say something like…Tools the agent uses
“Are these generated molecules novel vs my training set?”audit_generated_library
“Give me ADME and tox for this one SMILES.”adme_tox
“Find the best PDB structure for JAK2 and download it.”protein_select → fetch_pdb
“Prepare these ligands in 3D at pH 7.4.”prepare_ligands
“Define the pocket from the co-crystal ligand.”detect_pocket
“Dock my ligands to this target.”dock_to_target
“Rank the hits — it's a CNS target, weight BBB.”select_top_hits
“Plan the wet-lab validation experiments.”wet_lab_plan
“Search the literature for this target.”literature_search
“Write the final report as a PDF.”write_report

Every run saves outputs under runs/ with a run_manifest.json. Missing external engines (Open Babel, Vina, OpenMM…) are reported clearly instead of guessed around — see How it works.

Ask your agent. Get real science back.

Start with a free 3-day trial — no card required.

Start free trial See how it works