Loading ToolBench
A lightweight desktop supervisor for Serena MCP Server
Serena Desktop exposes 10 tools with minimal documentation. Tool names follow verb_noun convention (media_read_image, workspace_list, source_create_text_file), which is good. However, 4 of 10 tools (media_read_image, workspace_list, workspace_get, codegraph_explore) have NO visible input schemas in the provided source. Descriptions are present but generic (10-50 chars), below the 50-200 char LLM-optimized baseline. Write operations (source_*) include SHA-256 versioning for conflict detection, which is excellent for idempotency, but lack error recovery guidance. No tool annotations (readOnlyHint, destructiveHint) despite clear risk stratification. Output schemas are undocumented.
Explore code graph
Read and return image data
Create a new text file in workspace
Delete lines from a text file
Insert lines into a text file
Replace content in a text file using pattern matching
Replace lines in a text file
Write content to an existing text file
Four read-only tools (media_read_image, workspace_list, workspace_get, codegraph_explore) have no visible input schemas in source code.
All tool descriptions are under 50 characters (e.g., 'Read and return image data', 'List available workspaces for Discovery'). LLM-optimized descriptions should be 50-200 chars and include WHEN to use the tool and what it returns.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields without knowing response structure. source_* tools return success/failure but field names and types are undocumented.
Inferred effective spec: 2025-06-18+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 50 | 2025-06-18+ | v2 |
Get workspace details for Discovery
List available workspaces for Discovery
Write operations lack error recovery guidance. If source_write_text_file fails due to SHA-256 mismatch (concurrent edit), the error message should guide the LLM to retry with updated hash or call a read tool to refresh state.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk stratification: 4 READ_ONLY tools vs 6 WRITE tools. Annotations enable agents to reason about safety and retry logic.