Single tool with acceptable naming but significant gaps in parameter descriptions and output documentation. The tool name 'create_container' follows verb_noun convention correctly. Input schema is present with type definitions for all three parameters. However, parameter descriptions are embedded in the tool description rather than provided as individual field annotations in the schema object, making them harder for LLMs to parse. Output schema is completely undocumented, the tool returns a string but there is no specification of what that string contains or what structure the caller should expect. Security input validation (_safe_token) is well-implemented server-side, mitigating command injection risks. Error handling is present but generic (returns error strings without recovery guidance). The tool's risk classification as WRITE is correct but not leveraged for confirmation or dry-run patterns.

• Add structured output documentation: Document return value as JSON object with fields like {"container_id": string, "status": string, "message": string} instead of generic string return.
• Move parameter descriptions into JSON Schema 'description' fields: Each of image, container_name, dependencies should have its own 'description' property in the schema, separate from the tool description. This enables LLM parameter help and schema-driven validation.
• Add error recovery guidance: Instead of 'Error: image not found', return 'Image "foo" not found. Verify image exists with: docker images. Common images: ubuntu:latest, node:16, python:3.9-slim'.
• Implement dry-run mode: Add optional boolean parameter 'dry_run' (default false) that plans the operation without executing, or add a 'confirm_before_execute' tool annotation so clients can prompt for confirmation before container creation.
• Add tool annotations in fastmcp: Use @mcp.tool() decorator to register destructiveHint=true on create_container so clients know this modifies system state.
• Document idempotency: Clarify whether calling create_container twice with the same container_name will fail (non-idempotent, current behavior) or return existing container (idempotent). Add guidance: 'If container exists, tool returns error. Use get_container_status() to check first.'
• Specify timeout and retry behavior: Document that Docker operations may take seconds; specify timeout and whether LLM should retry on transient Docker daemon errors.

Inferred effective spec: <=2025-11-25.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry