Loading ToolBench

Reading the current corpus…

cxtools MCP server · D · ToolBench

← Browse

cxtools

github.com/mkusaka/cxtools

MCP server exposing Codex tools by calling the codex-rs crates in-process.

2025-06-18+STDIO4 tools

56/100

Definition Quality71

Protocol Readiness50

Supportability35

Spec posture

Inferred effective spec 2025-06-18+·current baseline 2026-07-28

Static source inference · medium confidence · evidence: structured output

No deprecated protocol patterns detected

Summary

Four tools with complete input schemas and descriptions. shell_command, apply_patch, and view_image have clear, actionable descriptions (60-120 chars). Schemas are well-typed with nullable fields documented. However, subagent lacks output schema documentation, and error handling guidance is minimal. No tool annotations (readOnlyHint/destructiveHint) despite shell_command and apply_patch being irreversible/write operations. Parameter descriptions are present but terse; no enum constraints or format specifications. Tests confirm structured output for shell_command but not for others.

Tools (4)

apply_patchwritesource verified73/100

The `apply_patch` tool can be used to edit files. Provide the patch in the apply_patch format (*** Begin Patch / *** Add File: / *** Update File: / *** Delete File: / *** End Patch).

shell_commandirreversiblesource verified80/100

Runs a shell command and returns its output. - Always set the `workdir` param when using the shell_command function. Do not use `cd` unless absolutely necessary.

subagentirreversiblesource verified57/100

Delegate a task to a non-interactive Codex agent.

view_imageread onlysource verified72/100

View a local image file from the filesystem when visual inspection is needed. Use this for images already available on disk.

Issues (5)

HIGHtool ↗

subagent tool lacks output schema documentation. Tests do not verify structured output; schema is inferred but not explicitly registered.

subagent

HIGHtool-annotation ↗

No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite shell_command and apply_patch being irreversible operations. LLMs cannot infer safety properties.

shell_commandapply_patchsubagent

MEDIUMrecovery-guide ↗

Error handling lacks recovery guidance. shell_command reports 'exit code 3' and 'timed out' but does not suggest next steps (retry, adjust timeout, check command syntax).

shell_commandapply_patch

MEDIUMtool-description ↗

Parameter descriptions are minimal (10-30 chars). 'Working directory for the command' and 'Shell script to run' lack format/constraint details. No guidance on when to use workdir vs cd.

shell_commandapply_patchsubagent

Recommendations

Add destructiveHint: true and idempotentHint: false to shell_command and apply_patch tool definitions to signal irreversibility to LLM clients.
Expand shell_command description to include error recovery: 'If exit code is non-zero, check command syntax and permissions. Timeout errors indicate the command took >timeout_ms; increase timeout or optimize the command.'
Document apply_patch format inline: 'Patch format: *** Begin Patch / *** Add File: <path> / +<content> / *** Update File: <path> / -<old> / +<new> / *** Delete File: <path> / *** End Patch'
Add output schema documentation for subagent tool. Define expected response structure (success/failure, agent output, execution time) and register via schema_for_output.
Expand parameter descriptions with constraints: timeout_ms should state '1 - 300000 ms (default 10000)', workdir should clarify 'Absolute or relative to server cwd; defaults to server cwd if omitted'.

Score history

Overall score trend

First recorded score · v2 rubric

56/100

Scored	Grade	Overall	Spec posture	Rubric
2026-09-23	D	56	2025-06-18+	v2