Visual MCP tools for exploring Convex databases - schema browser, component browser, dashboard, ER diagrams. Works as CLI, MCP server, or Claude Code plugin.
This server has critical definition quality gaps. While tool names follow verb_noun convention reasonably well (schema_browser, dashboard_view, etc.), the input schemas are severely under-specified. Most tools have 1-3 parameters with type declarations, but parameter descriptions are either missing or generic. For example, 'viewMode' in schema_browser has a description 'View mode: list, graph, or components' but provides NO enum constraint to prevent hallucinated values. The dashboard_view tool accepts arrays for 'metrics' and 'charts' but with no item type or structure documentation. None of the tools document their output schemas, making it impossible for LLMs to plan downstream actions or extract chaining IDs. Error handling is not visible in the code samples provided. The 'codebase_subway_map' tool accepts a 'root' directory parameter with no validation guidance or sanitization hints, raising path traversal concerns. Overall, these tools are defined at the bare minimum level: names and parameter names exist, but LLM-usable constraint documentation is absent.
Visual map of codebase file dependencies
Browse installed Convex components and their schemas
Live charts and metrics for Convex data
Kanban view of scheduled functions and agents
Interactive UI for browsing Convex database schemas
Mermaid ER diagram of database relationships
Compare declared vs inferred schema
Heatmap showing table write activity
Missing enum constraints on string parameters. 'viewMode' in schema_browser accepts 'list', 'graph', or 'components' but is declared as a free-form string type. LLMs will hallucinate invalid modes like 'table' or 'visual'. Parameters must declare enums explicitly or validation will fail silently.
No output schema documentation. Tools like 'dashboard_view', 'schema_diagram', and 'codebase_subway_map' do not document what fields they return. LLMs cannot plan downstream actions, extract chaining IDs, or know whether a result succeeded without documented output types.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 38 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 43 | - | v1 |
Analysis of write conflicts from logs
schema_drift tool has ZERO input parameters defined but is callable. The input schema is empty ({}), making it unclear whether the tool reads from environment, files, or accepts parameters. This violates the principle of explicit interface definition.
Array parameters lack item type specifications. 'dashboard_view' accepts 'metrics' and 'charts' as arrays but does not declare what the array items are (objects? strings? what fields?). LLMs cannot construct valid payloads without item schemas.
Path traversal vulnerability in 'codebase_subway_map'. The 'root' parameter accepts file system paths with no validation or sanitization guidance. An LLM could be prompted to pass '../../../etc/passwd' or similar. Tool description must state expected directory constraints and implementation must validate.
No error handling or recovery guidance visible. If a tool fails (e.g., missing Convex credentials, invalid table name), LLMs have no instruction on what to do next. Errors should guide recovery, not deadlock the agent.
Generic parameter descriptions. Many parameter descriptions like 'Metrics to display' or 'Charts to render' are too vague for LLMs to reason about valid options. Descriptions must state the expected format, structure, and constraints inline (e.g., 'Metrics: comma-separated list of metric names (cpu, memory, latency, error_rate)').
No pagination or result limit documentation. Tools returning lists (schema_browser, component_browser) do not document max result counts or pagination. If these return hundreds of items, LLM context will be exhausted.
Missing descriptions on boolean parameters. 'showInferred' in schema_browser says 'Show inferred schema fields' but does not clarify: What happens if true? What happens if false? What are the performance implications? LLMs need explicit guidance.