Loading ToolBench
Portable Windows developer MCP server with an exact 49-tool contract. Gives any AI a complete Rust + Windows dev shell.
Programmer-Wander provides 18 tools with mostly complete schemas and descriptions. Naming follows verb_noun conventions (read_file, write_file, cmd, powershell). Descriptions are present and substantive (avg ~120 chars), explaining WHAT each tool does. However, parameter descriptions are inconsistent, many lack detail on constraints, formats, or valid ranges. Error handling guidance is minimal; most tools lack recovery hints. Output schemas are not documented. Several tools expose dangerous operations (kill_process, cmd, powershell) without confirmation patterns or permission gates. The server is Windows-only and STDIO-based, limiting remote accessibility.
Execute a Windows Command Prompt command through cmd.exe /C and return output with exit code.
Copy file with metadata preservation.
Create directory recursively.
Guarded single-file text replacement: exact literal match, fails unless the occurrence count equals expected_replacements. Use transform_find_replace for bulk/regex sweeps.
Kill process by PID.
List directory contents recursively.
List running processes.
Destructive tools (kill_process, cmd, powershell) lack confirmation patterns or dry-run support. Agents can execute irreversible operations without safeguards.
Output schemas are not documented. LLMs cannot predict response structure, forcing them to guess at field names and types for downstream tool chaining.
Parameter descriptions lack constraint details. E.g., 'timeout' has no min/max; 'max_kb' default is 100 but no upper bound stated; 'depth' in list_dir has no range.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | <=2025-11-25 | v2 |
Real long-lived shell process with in-memory state (cwd, env, REPL variables). Actions: create, run, list, destroy, cd, env, history, read. Unlike shell_session (which persists to disk and spawns fresh processes), live_shell keeps a single process alive across calls - essential for interactive tools like Python REPL, Node REPL, or psql.
Move or rename file.
Execute PowerShell command. Most versatile single tool for Windows.
Check if a Rust project is ready for deployment (Cargo.toml, src/ directory, etc.)
Read file with smart options: search for pattern, range of lines, tail with delta polling, or auto-truncate large files.
Make HTTP request. With save=<path>, streams the body to disk instead of returning it (absorbs the former http_download, including resume-via-Range).
Check health and memory usage of running servers
State session: remembered cwd + env applied to each fresh command. Actions: create, run, list, destroy, cd, env (set/get), history, read (recent outputs). State auto-persists to disk on every change - survives crashes and restarts with no checkpoint ceremony. For an interactive live process (REPL, in-memory shell variables), use live_shell instead.
Run saved shortcuts and/or raw command sequences in a session. mode=list shows saved shortcuts; mode=run executes names=[] (saved, with params substitution) and/or commands=[] (raw), stopping on first error by default.
Validate MCP configuration file
Write content to file. Creates parent directories if needed. mode=append streams to the end of the file.
Error handling lacks recovery guidance. Tools return exit codes or generic errors without actionable next steps for LLMs (e.g., 'file not found, try list_dir first').
Simple file operations (copy_file, move_file, create_dir) have minimal descriptions (<50 chars). LLMs cannot distinguish when to use each vs alternatives.