Loading ToolBench

Reading the current corpus…

ocular MCP server · B · ToolBench

ocular

github.com/xyun1996/ocular

MCP server that gives coding agents vision through OpenAI-compatible multimodal models, with reliable binary image uploads for remote deployments.

2026-07-28+STDIO8 tools

70/100

Definition Quality78

Protocol Readiness50

SupportabilityNot measured-

Spec posture

Inferred effective spec 2026-07-28+·current baseline 2026-07-28

Static source inference · medium confidence · evidence: stateless requests

Current-spec patterns detected

Summary

Ocular provides 8 vision-focused tools with complete input schemas and descriptions. All tools follow verb_noun naming (analyze_*, extract_*, create_*) and include parameter descriptions. However, output schemas are not documented, responses are returned as JSON text without formal schema declarations. Error handling is minimal; tools return generic error objects without recovery guidance. The create_upload_session tool has an unusually long description (400+ chars) that could be condensed. Most tools are read-only (good for safety) but lack explicit readOnlyHint annotations in the protocol.

Tools (8)

analyze_chart_imageread onlyauthsource verified80/100

Analyze a chart image and return labels, trends, approximate values, and limitations.

analyze_imageread onlyauthsource verified80/100

Analyze a general image and return structured JSON for a coding agent.

analyze_ui_screenshotread onlyauthsource verified82/100

Analyze a webpage, app, UI, or design mockup screenshot for frontend implementation work.

compare_ui_screenshotsread onlyauthsource verified77/100

Compare two UI screenshots and return implementation-useful visual differences.

create_upload_sessionread onlyauthsource verified65/100

Returns upload instructions for the binary side channel (use this to analyze LOCAL image files without base64 corruption). Before uploading, verify the local file exists and is non-empty with Bash (e.g. `ls -la <path>` or `test -f <path> && stat -c %s <path>`). Then PUT the raw bytes to the returned upload_url with curl --data-binary (NOT base64); the response gives a file_id to pass to any vision tool. Stateless -- call once to learn the endpoint.

diagnose_error_screenshotread onlyauth

Issues (5)

HIGHtool ↗

Output schemas not documented. Tools return JSON text without formal schema declarations. LLMs cannot plan downstream operations or validate response structure.

analyze_imageextract_text_from_imageanalyze_ui_screenshotdiagnose_error_screenshotcompare_ui_screenshotsextract_table_from_imageanalyze_chart_imagecreate_upload_session

HIGHrecovery-guide ↗

No error recovery guidance. Tools return generic error objects (toErrorResult) without actionable next steps. LLMs cannot self-correct or know whether to retry.

analyze_imageextract_text_from_imageanalyze_ui_screenshotdiagnose_error_screenshotcompare_ui_screenshotsextract_table_from_imageanalyze_chart_imagecreate_upload_session

MEDIUMtool-description ↗

create_upload_session description is 400+ chars (exceeds 200-char baseline). Verbose instructions about curl and file verification should be condensed or moved to documentation.

Recommendations

Document output schemas for all tools. Define the JSON structure returned (e.g., 'Returns {analysis: string, confidence: number, recommendations: string[]}') so LLMs can plan chaining.
Add readOnlyHint: true to all tool definitions in mcp-server.ts. This signals to agents that these tools are safe to call without side effects.
Enhance error handling in tool-runner.ts. Replace generic toErrorResult with categorized errors: 'Image validation failed: file exceeds 20 MB. Try a smaller image.' Include the constraint violated and the invalid value.
Condense create_upload_session description to ~150 chars. Move detailed curl/bash examples to a README or inline comments, not the tool description.
Add array schema validation for compare_ui_screenshots.images. Document: 'Array of 2-4 image objects, each with image_base64 (string) and mime_type (string).' Validate length in the tool handler.

Score history

Overall score trend

First recorded score · v2 rubric

70/100

Scored	Grade	Overall	Spec posture	Rubric
2026-09-23	B	70	2026-07-28+	v2