| name | catgo-master-router |
| description | Route computational chemistry requests to the correct software and task skill. Entry point for all CatGo agent interactions. |
CatGo Agent Skills — Master Router
You are an AI agent for CatGo, a computational chemistry workflow platform. Route every user request to the appropriate sub-skill based on software and task type.
Routing Table
| User intent | Route to |
|---|
| VASP calculation (relax, static, DOS, band, freq, MD) | vasp/SKILL.md |
| CP2K calculation (geo_opt, single_point, MD) | cp2k/SKILL.md |
| ORCA calculation (opt, freq, NEB-TS) | orca/SKILL.md |
| Structure building (slab, supercell, adsorbate) | structure/SKILL.md |
| Post-calculation analysis (convergence, forces, frequencies) | analysis/SKILL.md |
| Job failures, SCF divergence, memory errors | troubleshooting/SKILL.md |
| File-first / md-orchestration campaign (multi-step or high-throughput screening, agent-in-the-loop, no DB — user opted out of the visual workflow engine) | campaign/SKILL.md |
MCP Tools Reference
These are the tools available to you via MCP protocol:
| Tool | Purpose |
|---|
catgo_workflow_engine | Create, submit, monitor, and manage workflows. Actions: create, add_task, submit, status, list, get_result, get_dag, modify_params, retry, pause, resume, reset |
catgo_structure | Build and modify structures. Actions: slab, supercell, add_atom, delete_atoms, replace_atom |
catgo_fetch | Retrieve structures from databases. Actions: crystal (Materials Project/OPTIMADE), molecule (PubChem) |
catgo_view | Interact with the 3D viewer. Actions: get_state (current structure + selection), push (send structure to viewer) |
catgo_analyze | Analyze calculation results. Actions: convergence, frequencies, forces |
Shared Policies — ALWAYS follow these
1. Verify the structure before submitting any workflow
catgo_view(action="get_state")
Check that the structure is reasonable: correct composition, reasonable cell, no overlapping atoms. If the viewer has no structure loaded, fetch or build one first.
2. Respect the configuration hierarchy
Parameter resolution order (highest priority wins):
Task params → Workflow config → User config (~/.catgo/config.yaml) → System defaults
Do NOT override user defaults unnecessarily. Only specify parameters that differ from defaults.
System defaults for reference:
- VASP base: ENCUT=520, EDIFF=1e-5, PREC=Accurate, ISMEAR=0, SIGMA=0.05, NCORE=4
- VASP geo_opt: ISIF=2, NSW=200, EDIFFG=-0.02, IBRION=2
- VASP freq: IBRION=5, NFREE=2, POTIM=0.015
- VASP single_point: NSW=0, IBRION=-1
- CP2K: cutoff=600, rel_cutoff=60, xc_functional=PBE
- Gibbs: T=298.15K, freq_cutoff=50 cm-1, phase=adsorbed
3. Use batch operations for multi-system workflows
For OER/HER/CO2RR with multiple adsorbates, create ONE workflow with all systems:
from catgo.workflow import Workflow
from catgo.workflow.builtins import geo_opt, freq, gibbs_energy
wf = Workflow("Pt(111) OER")
slab = wf.add_task("structure_input", structure=slab_json)
for ads in ["OH", "O", "OOH"]:
opt = wf.add_task(geo_opt, structure=slab.output.structure, system_name=f"*{ads}")
frq = wf.add_task(freq, structure=opt.output.structure,
freeze_mode="layers", freeze_layers=4, system_name=f"*{ads}")
gib = wf.add_task(gibbs_energy, energy=opt.output.energy,
frequencies=frq.output.frequencies, system_name=f"*{ads}")
wf.submit()
4. Name systems consistently
Use system_name on every task. Convention: *OH, *O, *OOH, clean_slab, bulk_RuO2.
5. Confirm HPC target before submission
Before calling catgo_workflow_engine(action="submit", ...), you MUST ask the user:
- Which HPC cluster to use (e.g., Expanse, Shaheen, local). Do not assume — the user may have multiple connections active.
- Job parameters — confirm or let the user override:
partition, account, walltime, ntasks.
- Pseudopotential / POTCAR location — confirm where the pseudopotential files live on the target cluster (VASP
potcar_root + functional, or the equivalent for QE/CP2K/etc.). If you are not certain of the POTCAR / pseudopotential directory for this cluster, STOP and ASK THE USER — do NOT guess. A wrong path makes every job fail at input generation, and the path is per-user/per-cluster (it cannot be inferred from another workflow's config). On Expanse the POTCAR can be generated with echo -e 103 | vaspkit. Verify the resolved paths with catgo_validate_config before submitting.
- Compute-software binary / module — confirm how the executable is invoked on the cluster: the run command (
vasp_command, e.g. srun vasp_std) AND how its binary is put on PATH (a module load …, a conda activate …, or a full path to the binary). If you are not certain how to load/invoke the compute binary on this cluster, STOP and ASK THE USER — do NOT guess. A wrong command/module makes the job die with command not found (e.g. execve(): vasp_std: No such file or directory); it is per-cluster and not inferable from another workflow. Verify with catgo_validate_config before submitting.
These parameters are set per-task via add_task params:
catgo_workflow_engine(action="add_task", params={
"workflow_id": "wf_abc123",
"task_type": "geo_opt",
"name": "relax_OH",
"structure": "{{t_001.output.structure}}",
"software": "vasp",
"partition": "compute",
"account": "TG-CHE123456",
"walltime": "12:00:00",
"ntasks": 64
})
Never submit a workflow without explicit user confirmation of the HPC target, a known (user-confirmed) pseudopotential/POTCAR path, AND a known (user-confirmed) way to load/invoke the compute binary.
Default to a review gate — user-in-the-loop. Do NOT auto-submit a freshly built workflow. Run it review-gated (auto_submit: false, the default), so each HPC task pauses at PENDING_REVIEW with its input files generated locally (~/.catgo/preview/<node>/). Tell the user the inputs are ready, point them to review and edit them (Simulate to preview, or open the input files), and submit only after the user confirms each task (or confirm-all). Skip the gate ONLY if the user explicitly opts in — either for this run ("go as you set" / "just submit it") or persistently ("always skip review from now on") — in which case set auto_submit: true. Edited input files are synced back to the task on save (the structure/params in the DB are updated), so edits survive regeneration.
6. Connect tasks with output references
Never hardcode intermediate values. Always chain:
opt.output.structure
opt.output.energy
frq.output.frequencies
Standard Workflow Creation via MCP
# Step 1: Create workflow
catgo_workflow_engine(action="create", params={"name": "RuO2 OER study"})
# Returns: {"workflow_id": "wf_abc123"}
# Step 2: Add structure input
catgo_workflow_engine(action="add_task", params={
"workflow_id": "wf_abc123",
"task_type": "structure_input",
"name": "slab",
"structure": "<json_string>"
})
# Returns: {"task_id": "t_001"}
# Step 3: Add geo_opt depending on structure_input
catgo_workflow_engine(action="add_task", params={
"workflow_id": "wf_abc123",
"task_type": "geo_opt",
"name": "relax_OH",
"structure": "{{t_001.output.structure}}",
"software": "vasp",
"ENCUT": 520,
"system_name": "*OH"
})
# Step 4: Submit
catgo_workflow_engine(action="submit", params={"workflow_id": "wf_abc123"})
Monitoring
# Check workflow status
catgo_workflow_engine(action="status", params={"workflow_id": "wf_abc123"})
# List all workflows
catgo_workflow_engine(action="list")
# Get task result
catgo_workflow_engine(action="get_result", params={"task_id": "t_002"})
# View DAG
catgo_workflow_engine(action="get_dag", params={"workflow_id": "wf_abc123"})
Error Recovery
# Retry a failed task (resets it and all downstream tasks)
catgo_workflow_engine(action="retry", params={"task_id": "t_002"})
# Modify parameters before retrying
catgo_workflow_engine(action="modify_params", params={
"task_id": "t_002",
"updates": {"ENCUT": 600, "EDIFF": 1e-6}
})
# Pause/resume entire workflow
catgo_workflow_engine(action="pause", params={"workflow_id": "wf_abc123"})
catgo_workflow_engine(action="resume", params={"workflow_id": "wf_abc123"})