Loading ToolBench
AI-powered project management tool with specialized agents for PRD creation and JIRA task management using Claude Agent SDK
This MCP server has critical definition quality issues that prevent safe production use. While it exposes 13 tools, most lack proper schemas, descriptions are minimal or missing, and there are severe security and composition problems. Tool naming is inconsistent (mcp__atlassian__* prefix is non-standard), parameters often lack type information, and error handling is absent. The server mixes file operations, user interaction, agent delegation, and Atlassian API access in ways that violate single-responsibility and composition patterns. Only 3 tools have both description AND input schema with typed parameters. Atlassian tools expose wildcard access and delegation without permission gates.
User interaction for asking clarification questions
File globbing for codebase exploration
Search for patterns within files
File operations for reading files from the codebase
Subagent delegation tool for using specialist agents
Create and write files
Wildcard access to all Atlassian MCP tools including Confluence operations
Wildcard tool access (mcp__atlassian__*) with no schema, no description, and no permission gates. This allows unrestricted access to all Atlassian resources including destructive Confluence operations without any validation or audit trail.
Task tool (subagent delegation) has no input schema defined. Parameters 'agent' and 'task' lack type definitions, constraints, or validation rules. This prevents the LLM from understanding valid agent names or task formats, and invites hallucinated agent names.
Six Atlassian tools (getAccessibleAtlassianResources, getJiraProjects, getJiraIssue, searchJiraIssues, updateJiraIssue, and the wildcard) have no visible input schema. Cannot verify parameter types, constraints, or required vs optional fields. This violates the schema documentation baseline.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 42 | <=2025-11-25 | v2 |
Create an epic, story, or task in Jira
Get cloud ID and available Atlassian resources
Get details of a specific Jira issue
List available Jira projects
Search for existing issues in Jira
Update an existing Jira issue
Tool names violate verb_noun convention. mcp__atlassian__* prefix is non-standard and does not start with an action verb (e.g., 'get_jira_projects', 'create_jira_issue'). The prefix obscures intent and makes it harder for LLMs to disambiguate tools. LLMs struggle with unprefixed, non-verb-starting names.
AskUserQuestion tool has vague name and minimal description ('User interaction for asking clarification questions'). Does not explain when to use it vs other tools, what the response format is, or how the user's answer is returned. LLM cannot determine proper invocation context.
No error handling documented for any tool. Tools provide no guidance on what to do if a Jira API call fails, file read fails, or user question times out. Errors lack categorization (retryable vs user-fixable vs fatal). Missing recovery guidance like 'Try search_users() first if you only have a name.'
No output schemas documented for any tool. LLM does not know what fields to expect from createJiraIssue (issue_id, issue_key, URL?) or searchJiraIssues (how many results? pagination? field names?). Undocumented output prevents downstream tool chaining and planning.
Glob, Read, and Write tools lack any description of input validation, path traversal protection, or security constraints. Write tool especially needs explicit permission gating and dry-run confirmation. No mention of which directories are safe to write to or read from.
createJiraIssue accepts a 'parent' parameter but no description of valid formats (issue key? issue ID? URL?). No constraint list of valid issue types. No indication of which fields are required vs optional. LLM cannot predict which fields to omit or how to reference parent issues.
searchJiraIssues expects a 'query' parameter described as 'JQL query to search for issues' but provides no examples, syntax hints, or link to JQL documentation. LLM must either guess JQL syntax or invoke the tool blindly. No pagination parameters visible (limit, offset, cursor).
Tool mixing and composition violation: Write, Read, Grep, and Glob tools handle filesystem operations, while Task handles agent delegation, while AskUserQuestion handles user interaction, while Atlassian tools handle external API access. No clear separation of concerns. An agent using these tools must juggle file I/O, human interaction, and third-party APIs in the same call context.