An MCP server that provides tools to interact with Amazon S3 buckets, enabling listing, reading, uploading, and deleting objects.
This S3 MCP server has functional tool definitions with proper schemas and basic descriptions, but falls short of production-grade quality. All 5 tools are explicitly registered with JSON Schema input definitions and descriptions. However, the descriptions are minimal (10 - 34 characters), well below the 50 - 200 character baseline for A+ tools. Parameter descriptions exist but are generic ('The name of the bucket', 'The key (path) of the object') and lack guidance on format, constraints, or when to use each tool vs. similar alternatives. No output schemas are documented, callers must infer structure from code. Error handling returns raw error.message without actionable recovery guidance. No tool annotations (destructiveHint on delete, readOnlyHint on read tools). The naming is clear (verb_noun pattern: list_buckets, read_object, put_object, delete_object) and follows conventions, but there is no guidance on mutual exclusivity, pagination limits, or prerequisites. This is a working D→C boundary server: functional for expert users but risky for autonomous agents.
Delete an object from an S3 bucket
List all S3 buckets
List objects in an S3 bucket
Upload an object to an S3 bucket
Read the content of an object from an S3 bucket
Descriptions are too short (10 - 34 chars) and generic. 'List all S3 buckets' lacks context on when to call this tool, what it returns, and how to chain it with other tools. Baseline is 50 - 200 chars for A+ tools.
No documented output schemas. Callers cannot see that list_buckets returns an array of {Name, CreationDate} objects or that list_objects returns {Contents: [{Key, Size, LastModified}]}. This forces LLMs to infer structure from code or trial and error.
Error handling returns raw error.message (e.g. 'Error: The specified bucket does not exist') with no recovery guidance. Per pattern:recovery-guide, errors should suggest next steps: 'Bucket not found. Call s3_list_buckets to see available buckets.'
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 46 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 44 | - | v1 |
No tool annotations for risk classification. delete_object and put_object should declare destructiveHint=true and writeHint=true respectively. read_object and list_* should declare readOnlyHint=true. This prevents agents from accidentally selecting write tools when a read is safer.
list_objects accepts maxKeys (default 1000) but no description of pagination. Per pattern:paginated-result, tools returning lists should document the limit, explain when it truncates, and offer cursor/offset+limit for continuation. Current code may silently truncate large result sets.
Parameter descriptions are minimal. 'The name of the bucket' tells an LLM nothing about format (DNS-compliant, 3 - 63 chars, lowercase). 'The key (path) of the object' doesn't explain if wildcards are allowed, if leading '/' is stripped, or case sensitivity. Add format and constraint details to every param.
No confirmation or dry-run support for destructive operations. delete_object is irreversible, if an agent passes the wrong key, data is lost. Pattern:confirmation-request suggests a dry_run param or a separate 'confirm_delete' tool for critical operations.
AWS credentials (S3_ACCESS_KEY_ID, S3_SECRET_ACCESS_KEY) are required via environment variables at startup. This is acceptable per pattern:secret-injection, but the server should validate that all three vars are present before starting. Current check logs an error and exits with code 1, which is correct. However, if running in a containerized environment, this failure mode may be unclear to operators.