Loading ToolBench
An MCP server that provides queryable, on-demand project context to LLMs
Jumble provides 13 tools with complete JSON schemas and clear descriptions. Tools are well-named with action verbs (list_, get_, store_) and serve a focused domain: queryable project context for LLMs. However, several parameter descriptions lack specificity about formats and constraints, and some tools expose implementation details that could be simplified. Output schemas are not explicitly documented in the visible code, which limits guidance for downstream tool composition. All tools are READ_ONLY except store_memory, which is appropriate for a context-serving tool. The tool set is cohesive and well-structured for its purpose.
Returns architectural info for a specific concept/area of a project, including relevant files and a summary.
Returns executable commands for a project (build, test, lint, run, dev, etc.)
Returns project-specific coding conventions and gotchas. Conventions are architectural patterns and standards; gotchas are common mistakes to avoid.
Returns a documentation index for a project, listing available docs with summaries. Optionally retrieves the path to a specific doc.
Retrieves a specific memory entry by key for a project.
Returns metadata about a specific project including description, language, version, entry points, and dependencies.
Parameter descriptions lack format/constraint specificity. 'project' is described only as 'The project name' without format (length limits, character restrictions, case sensitivity). 'query' and 'topic' similarly lack guidance on expected formats, which forces LLMs to guess valid inputs.
Output schemas are not documented in the visible code. Tools return formatted JSON (via format_* functions) but LLMs have no explicit schema documentation for response fields. This prevents agents from reliably extracting follow-up data and chains downstream tool calls.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 59 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 26 | - | v1 |
Finds files related to a concept or feature by searching through all defined concepts.
Retrieves a task-specific skill containing focused context and instructions for a particular task.
Returns workspace-level conventions and gotchas that apply across all projects in the workspace.
Returns a high-level overview of the entire workspace: workspace info, all projects with descriptions, and their dependency relationships. Call this first to understand the workspace structure.
Lists all projects with their descriptions. Use this to discover what projects exist in the workspace.
Lists available task-specific skills for a project. Skills provide focused context for specific tasks like adding endpoints, debugging, etc.
Stores a memory entry (key-value pair) for a project. AI agents can use this to persist learned information, preferences, or context over time.
Tools that return lists (list_projects, list_skills, get_docs, get_workspace_overview) lack pagination parameters (limit, offset, cursor). No indication of result caps, which could cause context window exhaustion if workspaces contain many projects or skills.
store_memory accepts an optional 'source' parameter for tracking agent identity, but no parameter descriptions explain what values are valid or expected. This could lead to inconsistent memory keys across different agent runs.
Error handling and recovery guidance are not visible in the tool definitions. No indication of which errors are retryable (e.g., malformed project name vs. network timeout) or how LLMs should recover (e.g., 'Project not found. Call list_projects() to see available options.').