MCP server providing tools for Globus Compute and Globus Transfer services
The Globus MCP server demonstrates solid foundational quality with well-structured tool definitions and comprehensive parameter descriptions. All 10 tools have explicit names following verb_noun conventions (globus_compute_*, globus_transfer_*), descriptions present, and JSON Schema input definitions with type constraints and field descriptions. However, output schemas are not documented in the visible code, only Pydantic response classes are referenced without their fields being shown. Parameter descriptions are generally good (averaging ~80 chars) but some tools lack actionable error guidance. The server achieves a C+ to B- grade: definitions are production-adjacent but missing documented output schemas and error recovery patterns that would push it to A-range. No critical security issues detected (no secrets in params, proper use of Context injection). Schema coverage is the primary gap.
Retrieve the status and result of a Globus Compute task.
List Globus Compute endpoints that the user has access to.
Register a Python function with Globus Compute. Use globus_compute_submit_task to run the registered Python function on an endpoint.
Register a shell command function with Globus Compute. Use globus_compute_submit_task to run the registered shell command on an endpoint.
Submit a function execution task to a Globus Compute endpoint. Use globus_compute_get_task_status to monitor progress and retrieve results.
Get a list of Globus Transfer task events to monitor the status and progress of a task. The events are ordered by time descending (newest first).
Output schemas not documented in visible code. Tools return Pydantic response classes (ComputeEndpoint, ComputeTask, ComputeSubmitResponse) but the actual field definitions and types are not shown in the provided source. LLMs cannot plan downstream operations or extract required fields without knowing response structure. This violates the pattern:tool requirement that 'Document the output schema.'
Tool name 'globus_transfer_list_endpoints_and_collections' violates the single-responsibility principle. The 'and' signals two distinct operations: listing endpoints and listing collections. This should be split into 'globus_transfer_list_endpoints' and 'globus_transfer_list_collections' so each tool has one clear, verifiable job.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | C | 63 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 0 | - | v1 |
List contents of a directory on a Globus Transfer collection
List Globus Transfer endpoints and collections that the user has access to, filtered based on the provided scope.
Use a filter string to search all Globus Transfer endpoints and collections that are visible to the user.
Submit a transfer task between two Globus Transfer collections. Use globus_transfer_get_task_events to monitor the task's progress.
Error handling returns generic ToolError exceptions (e.g., 'Failed to register Python function: {e}') without recovery guidance. Per the pattern:recovery-guide, errors must tell the LLM what to do next. E.g., 'Failed to register function, ensure the function_code is syntactically valid Python and imports only standard library or globus_sdk modules.'
No confirmation/dry-run pattern for destructive or write operations. globus_compute_register_python_function, globus_compute_register_shell_command, globus_compute_submit_task, and globus_transfer_submit_task modify state but offer no way to preview or confirm before execution. Agents can accidentally submit compute tasks or transfer large files. Consider adding a 'dry_run' parameter or a separate 'preview_*' endpoint.
Function_args parameter type is 'tuple' in schema but tool description says 'Positional arguments for the function.' Tuples and lists are semantifcally identical for JSON, but the parameter name and actual input schema should match. Consider renaming to 'function_args_list' or clarifying in the description that it accepts an array.
No declared permission scopes for tools. Tools like globus_compute_submit_task and globus_transfer_submit_task have real operational impact but do not declare what permissions/scopes they require (e.g., 'compute:submit', 'transfer:write'). This prevents least-privilege agent configurations and clear audit trails (pattern:scope-declaration).