Loading ToolBench
MCP server for workspace file reading and tool catalog management with support for both stdio and HTTP transports
Single tool 'rt_read_local_file' has excellent naming (verb_noun pattern), comprehensive description (234 chars, within 10-1024 baseline), and well-structured input schema with proper types and descriptions. Parameter descriptions explain format, constraints, and behavior clearly. Output schema is documented in description (startLine/endLine/totalLines/TRUNCATED). However, output schema is not formally declared in JSON Schema format, only described in prose. Error handling guidance is present but could be more actionable. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite READ_ONLY risk classification.
Read a file from the workspace, whole or as a line window. startLine is 1-based inclusive; omit lineCount to read to the end. Every result reports startLine/endLine/totalLines, so paging is a number, not a guess. A read too large for this server's caps comes back marked TRUNCATED, still carrying the true totalLines and the startLine to continue from.
Output schema not formally declared in JSON Schema. Description mentions startLine/endLine/totalLines/TRUNCATED but no structured output schema visible in tool registration.
Tool marked READ_ONLY risk but lacks tool annotations (readOnlyHint). Current spec (2026-07-28) supports readOnlyHint to signal safe-to-retry operations.
Error handling description is minimal. Does not guide LLM on recovery actions (e.g., 'If file not found, try listing directory with list_files tool').
No pagination or result-size limits documented. Large files could return unbounded output, risking context window exhaustion.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | <=2025-11-25 | v2 |