Loading ToolBench
Multi-layer agent-memory MCP server (working, episodic, semantic, procedural, affective) for Claude Code and any MCP client.
Dokoro is a STDIO-only MCP server with 16 tools spanning tagging, analysis, and asset management. Most tools have schemas visible in source (ai-tagging-tools.ts, analysis-tools.ts, asset-tools.ts) but suffer from inconsistent description quality, missing parameter-level descriptions in several tools, and weak error handling guidance. Naming is clear and verb-forward (dokoro_auto_tag, dokoro_validate_tags, dokoro_analyze_patterns), but many parameter descriptions are sparse or generic. No tool annotations (readOnlyHint/destructiveHint) despite clear risk levels. Archive sweep tool has destructive capability (default dryRun=false) with inadequate safeguards. Output schemas are largely undocumented in the source, we cannot verify what fields responses contain.
Find recurring issues, patterns, and technical debt across dokoros
Archive stale workspace files: daily/*.md older than olderThanDays move to archive/daily/<ISO week>/ (the current ISO week and files with live advisory claims are NEVER touched), and completed/validated plans older than planOlderThanDays move to .mcp/plans/archive/<YYYY-MM>/ (still listed by dokoro_plan_list, read-only). IMPORTANT: dryRun defaults to FALSE — by default files ARE moved; pass dryRun:true to preview without changing anything. status_only:true skips sweeping entirely and reports .mcp/archive-status.json from the last run.
Automatically suggest tags for dokoro content
Track progress on epics/features with burndown visualization
Visualize feature dependencies and blocking relationships
Get current feature implementation status
Output schemas undocumented: Tools return responses but source does not declare response field types, nullability, or pagination structure. LLMs cannot plan downstream tool chains or know what fields to extract.
Missing parameter-level descriptions: Tools like dokoro_validate_tags, dokoro_tag_stats, dokoro_feature_status, and dokoro_pending accept parameters with minimal or no description. Example: dokoro_validate_tags input 'tags' has description 'Tags to validate' (generic); dokoro_feature_status 'feature' only says 'Feature name to check status' without explaining expected format.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | 2026-07-28+ | v2 |
List all saved assets in the dokoro
Find stale or incomplete work items
Save any file to the dokoro assets folder with automatic organization.
Save an image to the dokoro assets folder. Supports copying from a path or saving base64 data.
Analyze and tag multiple dokoro entries
Get statistics about tag usage
Generate automated test suggestions based on feature and regression history
Generate chronological development history
Validate tags against the taxonomy
Track development productivity patterns and metrics
No tool annotations (readOnlyHint/destructiveHint/idempotentHint): Tools declare Risk levels (READ_ONLY, WRITE, DESTRUCTIVE) but source code does not register MCP tool annotations. dokoro_archive_sweep is DESTRUCTIVE (moves files by default, dryRun defaults to FALSE) yet lacks annotation warnings. LLMs cannot infer safety level from metadata.
Insufficient error guidance: Source shows no error recovery hints. Tools that fail (e.g., feature not found, pattern analysis timeout) do not suggest next steps. Pattern: 'Feature not found. Try dokoro_list_features()', source provides no such hint.
Destructive default dryRun=false: dokoro_archive_sweep defaults dryRun to FALSE, meaning files are moved by default without explicit confirmation. Source description warns this is 'IMPORTANT' but MCP client cannot confirm before execution. Risk of accidental data loss.
Enum values not enforced at registration: Tools like dokoro_dependency_graph accept format enum [mermaid, json, dot] and dokoro_timeline accept range enum [week, month, quarter, all], but source does not show validation rejection of invalid values (e.g., 'invalid_format'). LLM may pass unsupported values.
Missing pagination/result limit guidance: Tools returning lists (dokoro_analyze_patterns, dokoro_velocity_insights) do not document max result count or pagination support in descriptions. If analysis returns 100+ patterns, response balloons context window.