Static source inference · medium confidence · detected: Logging
Deprecated protocol patterns detected
Summary
This server has 35 tools but exhibits pervasive quality gaps across naming, descriptions, parameters, and schemas. Tool names are poorly chosen (mixing POST/PUT/GET HTTP verbs rather than semantic action verbs), descriptions are generic and often inaccurate (e.g., 'List of user items' for put_v1_tanssevents; 'Get a list of phone calls' for put_calls_v1), and parameter schemas are minimally documented. Input schemas are visible but lack type specificity, most show only generic 'object' types with vague descriptions like 'domain object to be saved' or 'filter settings' without field definitions. No output schemas are documented anywhere in the provided code. Error handling guidance is absent. The server conflates HTTP method semantics with tool semantics, violating the single-responsibility principle and making LLM tool selection unreliable.
Tool names use HTTP verbs (post_, put_, get_, delete_) rather than semantic action verbs (create_, update_, list_, search_). This violates naming conventions and makes LLM tool selection unreliable. Examples: post_v1_domains should be create_domain; put_v1_chats should be update_chats; put_v1_callbacks should be list_callbacks.
Descriptions are generic, inaccurate, or contradictory. Examples: 'List of user items' for put_v1_tanssevents (a PUT endpoint should update, not list); 'Get a list of phone calls' for put_calls_v1 (PUT is update, not retrieval); 'domain object to be saved' without specifying required/optional fields. LLMs cannot reliably infer when to use these tools.
put_v1_tansseventsput_calls_v1put_v1_callbacks
Recommendations
Rename all tools to use semantic action verbs, not HTTP methods. Create a mapping: post_v1_domains → create_domain, put_v1_domains → update_domain, get_v1_domains → get_domain, delete_v1_domains → delete_domain. Apply consistently across all 35 tools. Reference the verb_noun pattern (send_email, create_ticket, list_users) from production baselines.
Rewrite all descriptions to be LLM-optimized (50 - 200 characters, following the pattern: WHAT the tool does, WHEN to use it, any prerequisites). Example: 'Creates a new domain in the TANSS system. Use this to add infrastructure domains for employee assignment and phone routing.' Replace vague terms like 'object to be saved' with concrete actions.
Expand input schemas from generic 'object' types to fully specified JSON Schema. For each 'body' parameter, document required fields, field types (string, integer, boolean, etc.), and per-field descriptions. For list/filter operations, explicitly list which filters are available (e.g., status, created_after, employee_id).
Document output schemas for all tools. Provide a structured description of the response (e.g., 'Returns {id: integer, name: string, created_at: ISO8601, status: enum}') or include example JSON structures in tool metadata. This unblocks tool chaining and LLM reasoning.
Add tool annotations (readOnlyHint, destructiveHint) to signal risk. Mark delete_* and destructive operations with destructiveHint=true. Mark read-only operations with readOnlyHint=true. This helps MCP clients enforce least-privilege policies.
Spec posture evidence
Inferred effective spec: <=2025-11-25.
Relies on Logging (deprecated) - log to stderr or use OpenTelemetry
Input schemas are vague and incomplete. Most parameters are typed as generic 'object' (e.g., 'body': {'type': 'object', 'description': 'domain object to be saved'}) without specifying required fields, field types, or field descriptions. This forces LLMs to guess structure and invites invalid payloads.
No output schemas are documented for any tool. The code shows tool definitions but no CallToolResult response structures, return field documentation, or data structures agents can expect. This prevents downstream tool chaining and forces LLMs to blindly parse responses.
No error handling guidance. Tools lack descriptions of failure modes, recovery steps, or actionable error messages. For example, delete operations (delete_v1_domains, delete_v1_employeeassignment, delete_chats_participants) do not warn about irreversibility or suggest confirmation workflows.
Destructive operations lack safety patterns. delete_v1_domains, delete_v1_employeeassignment, and delete_chats_participants have no confirmation workflow, dry-run mode, or permission checks documented. Agents can delete critical data without safeguards.
Parameter descriptions are under-specified. Examples: 'id of the domain that shall be fetched' does not clarify data type, range, or format; 'Ids of the employees (comma separated)' is manual parsing that LLMs often get wrong; 'filter settings' is too vague. Detailed constraints should specify expected format, allowed values, and ranges.
Conflicting or ambiguous tool functionality due to HTTP verb naming. Multiple endpoints map to the same semantic action with different HTTP methods (e.g., put_v1_callbacks is 'Get a list' but uses PUT; put_calls_v1 is also a list operation). This creates confusion and risk of calling the wrong tool.
No tool annotations for risk classification. None of the tools have documented tool annotations (readOnlyHint, destructiveHint, idempotentHint) in the MCP spec sense, even though the risk levels are tracked in metadata. This prevents MCP clients from understanding which operations are reversible.
No pagination guidance for list operations. Tools like get_v1_availability, get_v1_callbacks, get_calls_v1, get_v1_chats, get_v1_employeeassignment do not document pagination parameters (page, limit, offset, cursor) or result count limits. Large datasets could exhaust context windows.
Implement confirmation workflows for destructive operations. For delete_v1_domains, delete_v1_employeeassignment, delete_chats_participants, add a dry_run parameter or a separate confirm_delete_{resource} step. Warn agents of irreversibility in the description.
Consolidate PUT-based list/filter operations under semantic list_ names. Replace put_v1_callbacks ('Get a list') with list_callbacks. Similarly rename put_calls_v1 → list_calls, put_v1_phonecalls → list_phonecalls, put_v1_chats → list_chats, put_v1_tanssevents → list_events. Update descriptions to match the new names.
For all list/query tools, document pagination: add limit (default 20, max 100), offset or page parameters, and return total_count in the response. Example: 'Returns up to 20 results. Use offset to page through larger datasets. Include total_count so agents know when to stop.'
Specify parameter types and constraints more precisely. Replace 'Ids of the employees (comma separated)' with 'Comma-separated list of employee IDs (integers). Example: 123,456,789'. For IDs, clarify range (positive integers) and any length constraints. For enums, use JSON Schema enum constraint rather than text examples.
Add error recovery guidance to all tools. Example for delete operations: 'Returns 404 if domain not found. Returns 403 if insufficient permissions. Returns 409 if domain has active assignments; remove assignments first.' This guides agent retries.
Eliminate generic descriptions for filter/settings parameters. Instead of 'filter settings', enumerate expected fields. Example: 'Filter object with optional fields: status (open|closed|pending), created_after (ISO8601), employee_id (integer).'
Document which operations are idempotent. E.g., marking all events as seen (post_all_seen) is likely idempotent; creating a callback (post_v1_callbacks) is not. This helps agents decide whether to retry on failure.
Add role/permission hints to sensitive operations. E.g., 'Requires admin role' or 'Requires write:domain permission' so agents understand access control and can fail gracefully if denied.