Loading ToolBench
Multi-tool MCP server for the MarvisX open-core agent-native company-brain platform. Provides tools for knowledge graph operations, search, project management, ingest triage, workspace file operations, and audit logging.
Critical gaps across all dimensions. Most tools (17/24) lack visible input schemas entirely, cannot verify parameter types or constraints. Descriptions are present but generic (avg ~40 chars, well below 194-char baseline). No evidence of output schema documentation. Tool names lack consistent verb-noun structure (e.g., 'handoffs', 'learnings', 'costs' are nouns, not actions). Workspace tools have minimal parameter descriptions. Ingest tools sourced from test files, not production definitions, tool registration unclear. No error handling guidance, no security annotations, no idempotent/destructive hints visible.
Approve a pending ingest item for processing
Classify an ingest item using the LLM classifier
Query or manage cost tracking data
Query the knowledge graph structure
Traverse the knowledge graph from a starting node
List or manage handoff points in the knowledge graph
Delete a node or edge from the knowledge graph
17 of 24 tools have NO visible input schema. Cannot verify parameter types, constraints, or required fields.
Noun-based tool names (handoffs, learnings, costs, monitoring) violate verb-noun convention. LLMs cannot infer action from name alone. Should be: list_handoffs, manage_learnings, query_costs, get_monitoring_data.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 41 | <=2025-11-25 | v2 |
Insert a new node or edge into the knowledge graph
Update an existing knowledge graph node or edge
CRUD operations on learning/insight nodes in the knowledge graph
List pending ingest items awaiting triage or approval
Access monitoring and observability data
Patch/update a pending ingest item
Reject a pending ingest item
Reparse an ingest item to extract structured data
Search the knowledge graph by semantic meaning or keyword
Generate a brief summary of the current session context
Upload a new ingest item for processing
Search for text patterns in workspace files
Read files from the workspace
Execute shell commands in the workspace (restricted)
List directory contents in the workspace
Write files to the workspace
Write haiku-format frontmatter metadata to an ingest item
All tool descriptions are generic and under 50 chars (baseline 194 chars). No context on WHEN to use, WHAT it returns, or dependencies. E.g., 'List pending ingest items awaiting triage or approval' lacks guidance on pagination, result limits, or output structure.
No output schema documentation visible for any tool. LLMs cannot plan downstream calls or extract required fields. E.g., does search() return result_count, next_cursor, or total_hits? Does kg_insert() return the new node_id?
Destructive tools (kg_delete, workspace_shell) lack confirmation/dry-run patterns and no error recovery guidance. No indication of retryability or side effects.