Loading ToolBench
Sandboxed file-access MCP server for Claude Desktop: loud failures, verified atomic writes, built-in undo, .docx reading
srcbridge demonstrates solid engineering with 19 well-defined tools covering file I/O, git operations, and document handling. All tools have descriptions (avg ~80 chars) and input schemas with typed parameters. Naming follows verb_noun convention (read_source, write_source, grep_sources). However, output schemas are not documented in the visible code, parameter descriptions are minimal (often single phrases), and error handling guidance is absent. The codebase shows production maturity (atomic writes, snapshots, path sandboxing) but the MCP interface lacks LLM-optimization patterns like recovery hints, actionable error messages, and structured output documentation.
Compute SHA1 checksums for files matching a pattern
Show unified diff of a file between two git commits
Apply multiple edits to a file in one operation with per-edit preview
Find and replace text in a file with anchor diagnosis, span support, and preview mode
Find files by name pattern with walk limit and announced skips
Show git status of a repository (staged, unstaged, untracked)
Search for text patterns across multiple .docx files
Output schemas not documented. No tool describes what fields it returns or their types. LLMs cannot plan downstream tool calls or extract structured data without knowing response shape.
Parameter descriptions are minimal (single phrases like 'Absolute path to the file to read'). Missing context on format constraints, valid ranges, and when to use each parameter. E.g., 'from_line' lacks min/max bounds or explanation of 1-indexing behavior.
No error handling guidance. Tools do not document what errors can occur, whether they are retryable, or what the LLM should do next. E.g., 'file not found' vs 'permission denied' vs 'path outside sandbox' all need distinct recovery paths.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | 2026-07-28+ | v2 |
Search for identifier/symbol references in source code
Search for text patterns in files with context lines and announced skips
List auto-snapshots for a file with timestamps and sizes
Show git commit history for a file with short oids
Read a file as base64-encoded bytes with size cap
Read a .docx file as Markdown with table/list structure preserved
Read a text file with optional line range, paging support, and announced truncation
Read a file at a specific git commit with history support
Read multiple files with aggregate byte budget and announced truncation
List directory contents recursively with depth limit and announced skips
Restore a file from its most recent auto-snapshot
Write or overwrite a text file with atomic write, verification, and auto-snapshot
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). LLMs cannot distinguish safe read-only tools from destructive ones without explicit hints. write_source and edit_source should be marked destructiveHint=true.
Pagination and result limits not documented. Tools like grep_sources and tree may return large result sets. No mention of limits, truncation behavior, or how to paginate. Agents may exhaust context windows.