FastMCP-compatible MCP Server for Cyclic Peptide computational tools. Provides synchronous tools for fast membrane permeability predictions and asynchronous job submission tools for long-running analyses and model training using Graph Convolutional Networks and Transformer models.
The server defines 8 tools with schemas and descriptions visible in the source code. However, there are significant quality gaps across naming, parameter documentation, output schemas, and error handling. Tool names follow verb conventions (predict_, submit_, get_, list_, cancel_) which is good. Descriptions are present but vary widely in quality and actionability. Most critical: no tool descriptions explain what the LLM should expect in return (output structure), parameter descriptions lack constraint details (ranges, formats, validation rules), and error handling provides no recovery guidance. The fastmcp framework wraps synchronous operations as async, but no evidence of structured error responses or per-request _meta handling. STDIO-only transport is a hard cap; this server cannot be used with hosted MCP clients.
Cancel a running cyclic peptide computation job. Args: job_id: The job ID to cancel Returns: Success or error message
Get execution logs from a job. Args: job_id: ID of the job to get logs for tail: Number of lines from end of log (optional) Returns: Dictionary with log lines
Get the results of a completed cyclic peptide computation job. Args: job_id: The job ID of a completed job Returns: Dictionary with the job results or error if not completed
Get the status of a submitted cyclic peptide computation job. Args: job_id: The job ID returned from a submit_* function Returns: Dictionary with job status, timestamps, and any errors
List all submitted cyclic peptide computation jobs. Args: status: Filter by status (pending, running, completed, failed, cancelled) Returns: List of jobs with their status
Predict PAMPA membrane permeability for cyclic peptides. Fast operation - returns results immediately (typically < 30 seconds). Args: input_file: Path to CSV file with SMILES column (for batch prediction) smiles: Single SMILES string for individual prediction sequence: Single amino acid sequence for prediction output_file: Path to save results CSV (optional for single inputs) config_file: Path to configuration JSON file (optional) Returns: Dictionary with prediction results. For batch: includes output file path. For single inputs: includes predicted PAMPA value.
No output schema documentation for any tool. Descriptions state what the tool returns ('Dictionary with...') but do not specify the structure, field names, or types. LLMs cannot plan downstream calls or extract data without knowing what fields to expect.
Parameter descriptions lack constraint information. Example: 'epochs' parameter has description 'Number of training epochs' but no valid range (e.g., 1-1000). 'learning_rate' lacks guidance on expected range (e.g., 0.00001-0.1). 'batch_size' has no constraint. LLMs will guess invalid values.
No error handling guidance in tool descriptions. A job may fail, be cancelled, or timeout, but the tool descriptions do not explain what errors can occur, which are retryable, or what the LLM should do next. 'get_job_result' returns 'error if not completed' but does not explain how to handle incomplete jobs.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 51 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 39 | - | v1 |
Submit machine learning model training job. This trains a GCN-Transformer model for membrane permeability prediction. Training time varies: 2-30 minutes for small datasets, hours for large ones. Args: input_file: CSV file with SMILES and target values output_dir: Directory for model outputs (weights, training history) epochs: Number of training epochs (default: 100) batch_size: Training batch size (default: 32) learning_rate: Learning rate (default: 0.001) test_split: Fraction for test set (default: 0.2) val_split: Fraction for validation set (default: 0.2) config_file: Path to configuration JSON file (optional) job_name: Optional name for tracking Returns: Dictionary with job_id for tracking the training job
Submit comprehensive analysis of cyclic peptide properties. This operation generates visualizations and statistical analysis. Typically takes 1-5 minutes depending on dataset size. Args: input_file: CSV file with SMILES and optional target column output_dir: Directory for analysis outputs (plots, reports) target_column: Target column for correlation analysis (default: "PAMPA") config_file: Path to configuration JSON file (optional) job_name: Optional name for tracking Returns: Dictionary with job_id for tracking. Use: - get_job_status(job_id) to check progress - get_job_result(job_id) to get results when completed - get_job_log(job_id) to see execution logs
Submission tools (submit_peptide_analysis, submit_model_training) lack duration expectations. Descriptions say 'Typically takes 1-5 minutes' and '2-30 minutes' but do not state SLAs, timeout handling, or how to monitor long-running jobs. LLMs need to know when to check status.
Tool definitions are registered and visible in fastmcp decorators, but no evidence of tool annotations (readOnlyHint, destructiveHint, idempotentHint) per MCP spec. submit_* tools are destructive (write state) but lack @destructive markers. Idempotent tools like get_* lack @idempotent markers.
Parameter 'tail' in get_job_log has type 'integer' but no bounds. LLMs may pass negative numbers, zero, or gigantic values. Description should state 'positive integer, 1-10000 (default: 100)' or similar.
list_jobs accepts status with enum constraint (pending, running, completed, failed, cancelled), good. But description does not explain what each status means or when transitions occur. An LLM may not understand that a 'pending' job will eventually become 'running'.
Job management workflow requires LLM to coordinate submit_* → poll get_job_status → get_job_result → optionally cancel_job. No tool batch operations for checking multiple jobs or cancelling multiple jobs. If 10 jobs are submitted, the LLM must call get_job_status 10 times.