bewerbungsfoto-check-mcp
Provides XING-targeted headshot quality checks and crop/resolution recommendations for profile photos, including static requirements and a checklist via the requirements tool.
Headshot Photo Check
A small English-language remote MCP server for checking the technical quality of professional profile and job-application photos. It turns measurable image issues into prioritized advice, without judging the person pictured. MIT licensed; no account, API key or paid analysis API.
Status: live; submission preparation only
The operator reports main v0.1.1 live at https://mcp.profilfoto-ki.de/mcp, via the Netlify proxy site mcp-profilfoto-proxy in front of the Cloudflare Worker. This branch prepares v0.1.2; it has not been deployed or submitted. The GitHub repository remains private. server.json proposes de.profilfoto-ki/headshot-check; no Registry publication was performed. See submission materials and remaining operator actions.
Related MCP server: GetMattered
What it checks
Two read-only tools:
Tool | Inputs | Outputs |
| Exactly one | Score 0–100, verdict, criteria with status/value/threshold/tip, per-target crop recommendations, prioritized tips, mode and coverage notes |
| Optional | Static recommendations, dimensions, formats and checklist |
Full mode checks resolution, aspect ratio, size/format, mean luminance and clipping, luminance contrast/dynamic range, overall color balance, experimental sharpness and border-background structure. It compares border and center edge metrics. Face size/framing returns not_checked; no unreliable face or skin-tone guessing is used. It performs no biometric identification or attractiveness, identity, age, gender, ethnicity or other personal-attribute judgments. It does not generate images, video or audio or return edited images; internal resizing supports analysis.
Reference logic and translated tips come from the existing read-only application checker. Port details document exact thresholds and discrepancies between its prose and code. LinkedIn/XING square crops and application portrait crops are recommendations, not official platform specifications. No official platform upload-size numbers are claimed. Added sharpness thresholds and service size limits are also recommendations/service limits, not platform rules.
A score averages checked criteria only (pass=100, warn=60, fail=0). Unchecked criteria do not affect it. Verdict severity follows the reference: any failure means technical issues, any warning means room for improvement. A high partial score does not mean an unchecked photo is suitable.
Examples
ChatGPT prompt: “Check my LinkedIn headshot at https://images.example.org/headshot.jpg and prioritize improvements.” Replace the illustrative URL with a public image you are authorized to use. The hostname is illustrative, not a supplied photo.
Claude prompt: “Check this base64 PNG for my application and explain which checks were skipped.” Supply the actual base64 as image_base64, with target: "application".
Illustrative output excerpt (not a measured person/photo):
{
"overall_score": 87,
"verdict": "Partial check: Usable with room for improvement",
"analysis_mode": "header-only",
"criteria": [
{"id":"resolution","label":"Resolution and print suitability","status":"warn","value":{"width":600,"height":800},"threshold":"pass: short >=1200, long >=1500; warn: short >=600, long >=750","tip":"Suitable for online applications and standard PDFs. For larger cover-page prints, use at least 1200 × 1600 pixels."}
]
}The actual result includes all nine criteria, coverage notes, crop recommendations and prioritized tips. To ask without uploading: “What crop and resolution should I prepare for XING?” calls the requirements tool.
Connect to the existing endpoint
All snippets below use the existing endpoint; this branch’s v0.1.2 changes require a later approved deployment. This is a stateless Streamable HTTP service with JSON responses and no authentication. Hosted assistants cannot reach your loopback development URL.
Claude: Settings → Connectors → Add custom connector → enter https://mcp.profilfoto-ki.de/mcp. Availability and menu names depend on account/workspace configuration. Claude Code:
claude mcp add --transport http headshot-photo-check https://mcp.profilfoto-ki.de/mcpChatGPT/Claude setup and verification notes are maintained in submission/README.md, with the current OpenAI submission guide. Account and workspace permissions affect access. Directory availability is not claimed.
Local development
npm install
npm run dev
# In another terminal:
npm run typecheck
npm test
bash scripts/smoke.sh http://127.0.0.1:8787
npm run secret-scanNo Cloudflare login, account bindings or secrets are needed for local development. Wrangler is a project devDependency. Default mode is auto with an offline Images binding. It resizes to at most 64 px before pixel analysis. No login is required. To exercise the diagnostic full decoder locally:
npx wrangler dev --port 8787 --ip 127.0.0.1 --var ANALYSIS_MODE:fullTests generate all JPEG/PNG fixtures in memory; no real-person images are committed. tests/benchmark.test.ts generates a 1500×2000 JPEG and prints wall/process CPU timings. No deploy commands are needed.
Protocol and limits
POST /mcp supports initialize, initialized notifications, tools/list, tools/call and ping. Versions: 2025-06-18, 2025-03-26, 2025-11-25; unknown initialize versions negotiate 2025-06-18. Unknown protocol headers return 400. Notifications return 202, GET /mcp returns helpful 405 JSON. GET /health returns {"status":"ok"}. CORS permits all origins and MCP headers. No session storage, SSE requirement or authentication.
Use image_url whenever possible. Base64 payloads above roughly 150 KB may exceed the Workers Free 10 ms CPU budget due to JSON parsing and decoding. This is practical guidance, not a schema limit: base64 remains supported up to 8,000,000 decoded bytes.
Image limit: 8,000,000 bytes; request-body limit: 10,670,000 bytes to allow base64 JSON. URL requests require HTTPS, public multi-label hostnames, image/* content type, an 8-second total timeout including body reads, at most 3 redirects and bounded body-read size enforcement. IP literals (including public IPs), credentials, custom ports, localhost and internal/reserved suffixes are rejected, including redirect destinations. DNS rebinding/private addresses behind public DNS names cannot be pinned by this fetch-based guard; production must add an egress/network policy or restricted image-host allowlist if stronger isolation is needed. Do not expose this as an internal-network proxy.
A best-effort per-isolate limiter permits 30 requests per minute per hashed network address, with a bounded transient map. It is not a global edge-wide quota; configure an appropriate Cloudflare rate rule before launch. No raw IP is stored or logged by application code.
Known limits
Auto mode uses the optional Images binding for JPEG/PNG/WebP pixel checks, with a maximum 64 px long side. Both URL and base64 inputs use the same binding transform over one bounded source buffer; URL reads preallocate from Content-Length when available and enforce the actual byte cap. Header parsing covers the complete JPEG/PNG/WebP buffer without source pixel decoding or an Images info() call. Header scores always describe the original file. Auto mode always tries the binding, including tiny images. Missing bindings, quota exhaustion and transform errors fall back to native full analysis for tiny JPEG/PNG (<=16,384 pixels), otherwise explicit analysis_mode: "header-only" with a reason. The checked-in binding works offline without login. Images Free includes 5,000 unique monthly transformations; new transformations beyond that fail without overage charges on Free. Operators must keep their account on Images Free. See sourced CPU/quota research.
Sharpness measures coarse blur only; reduced images conceal fine focus problems. No DC-only JPEG or AC-energy decoder is used (Option A selected). Full diagnostic mode is limited to 4 MP JPEG / 1 MP 8-bit PNG. PNG metadata/inflate guards remain. Original headers are bounded at 100 MP / 65,535 px per side. EXIF orientation/color handling may differ across resize implementations; re-export upright. Face framing, semantic subject detection and original-resolution focus are unchecked.
Local benchmark numbers and methodology are in BUILD-REPORT.md and machine-readable results. The previous full 1500×2000 JPEG decoder measured 238.51 ms wall / 343.02 ms process CPU. The new path avoids that full decode inside the Worker. Local HTTP wall times include the offline image service; local process CPU is not Cloudflare request CPU. The previous production pipeline measured 29–36 ms CPU for URL auto mode, 29–39 ms for base64 auto mode, 10–12 ms for URL header-only, and 16–17 ms for base64 header-only on a 351 KB JPEG. After the 64 px/lazy-decoder revision, operator production CPU was 49–79 ms for streaming JPEG URLs, 15 ms for binding-backed base64 of the 351 KB JPEG, and 49 ms for a tiny native PNG. The new buffered URL / binding-first revision has not been measured in production. Large JSON/base64 requests can still exceed the 10 ms Free budget; these measurements are no guarantee.
Privacy and operations
Photos and URLs are processed in memory and not stored by the application. Cloudflare runs the Worker and configured Images binding; Netlify provides the front proxy. URL fetching reaches third-party hosts selected by the user. Tool-call logs contain event, tool, success and rounded duration. Binding diagnostics include dimensions and bounded, defensively redacted errors. Rate-limit hashes have a 60-second validity window but can persist in idle isolate memory until another request or disposal. Provider retention is separate and must be confirmed by the operator. See the maintained privacy, terms and support texts. Operator details are filled from the profilfoto-ki.de Impressum; review before publication.
Optional Analytics Engine: after operator account/dataset setup, uncomment the [[analytics_engine_datasets]] block in wrangler.toml at the top level (outside [vars]), binding ANALYTICS, dataset headshot_calls. Calls are guarded by if/optional binding checks and send only blobs=[tool, ok], doubles=[ms]; console counters remain the fallback. Review the current Cloudflare Analytics Engine limits and free-tier allowance before enabling; no account or billing setup was performed here.
Before publication: confirm legal contacts/retention, publish the proposed HTTPS policy/support pages, confirm infrastructure retention and deployment CPU budget, prepare domain proofs, configure global rate/egress controls and an external /health uptime monitor, run the golden prompts in ChatGPT and Claude, record a demo video, and obtain Safak's approval for each deployment/submission. These are launch tasks, not completed actions. Review operational call analytics weekly for misrouting. Listing draft, changelog, build report.
Available Tools
2 toolscheck_headshot_photoCheck headshot photoARead-onlyIdempotent
Use this when the user wants technical checks of one professional profile or job-application photo from a public HTTPS URL or base64. Maximum 8 MB; image_url is recommended. Base64 above roughly 150 KB may exceed the Worker CPU budget. Auto mode resizes through an optional Images binding for pixel metrics, including WebP. Without it, only small JPEG/PNG inputs receive pixel checks; other inputs receive explicit header-only results. Sharpness detects coarse blur only. Face framing is not checked. URL fetching contacts user-selected external hosts and is limited to 8 seconds and 3 redirects. Cloudflare Images may internally resize the input for metrics; no edited image is returned. Operational call metadata is logged. Do not use for image creation, editing, personal-attribute judgments or identity verification. The image is analysed in memory and not stored. It does not generate images, video or audio or return edited images. It does not judge attractiveness, identity, age, gender, ethnicity or any other personal attribute; no face recognition or biometric identification.
Inputs: target, image_url, image_base64. Returns: notes, target, verdict, criteria, +5 more
get_headshot_photo_requirementsGet headshot requirementsARead-onlyIdempotent
Use this when the user wants a technical headshot checklist and crop recommendations for LinkedIn, XING or job applications. Returns static recommendations, not official platform upload limits; no image input or external network access. No persistent changes; operational call metadata is logged. Do not use for assessing an actual image or confirming platform acceptance. The image is analysed in memory and not stored. It does not generate images, video or audio or return edited images. It does not judge attractiveness, identity, age, gender, ethnicity or any other personal attribute; no face recognition or biometric identification.
Inputs: target. Returns: note, target, checklist, recommendations
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 2 tools
One tool analyzes an actual image, the other returns a static checklist, and each description explicitly rules out the other's use case ('Do not use for assessing an actual image'). There is no resource or action overlap, so misselection is essentially impossible.
Both names use snake_case with a verb prefix and the shared core noun 'headshot_photo' (check_headshot_photo, get_headshot_photo_requirements), which is predictable. The second name is noticeably longer and mixes an action verb with a noun-phrase suffix, but the pattern is still readable.
Two tools is thin for a server surface, though the underlying domain (analyzing one photo plus returning static guidance) is genuinely narrow. It sits at the borderline where the pairing covers the basics but leaves little room for anything else.
The pair covers the core loop: get a technical checklist, then check a real image against expectations, with clear documented limits (size, URL fetch, no face framing check). Gaps exist — no combined 'analyze and recommend fixes' output, no batch or multi-platform validation — but agents can work around them by combining the two tools.
Maintenance
Related MCP Connectors
Look up exactly what an upload form accepts (passport, visa and exam photo rules, each with its official source and the date it was checked), check whether a file will pass, and make links that fix files to fit, right in the person's browser. Free, no sign-in, and the server never receives files.
Technical checks and crop recommendations for professional profile and application photos.
YouTube and video analysis, C2PA marking for AI images, lossless EXIF/GPS removal. Pay per call.
Remove image backgrounds (transparent PNG/WebP) and look up passport photo rules.
Related MCP Servers
- AlicenseBqualityDmaintenanceDeterministic MCP server for job-landing pipeline that parses CVs into validated JSON Resume, detects sections, scores parse confidence, and exposes FR/US market profiles to help beat ATS and LLM screeners honestly.75MIT
- AlicenseAqualityBmaintenanceAuto-Apply That Ends on an Interview. Job matches, applications submitted on the employer's own hiring system automatically or after your approval, employer-receipt tracking, a resume rewritten for each role, cover letters and interview prep. Hosted remote server over Streamable HTTP with OAuth 2.1.412,455 npm4MIT
- AlicenseNot gradedqualityBmaintenanceEnables AI agents to run a deterministic job-application pipeline over MCP: scoring fit against a profile's evidence, drafting market-aware CVs and cover letters for 15 country conventions, and shipping documents only after an ATS-style gate checks page count, text layer, raster layout and naming. Writes require explicit confirmation, and every command returns a single JSON envelope with meaningful exit codes.1MIT
- AlicenseNot gradedqualityCmaintenanceEnables users to upload a resume to build a profile, discover jobs from company websites and official ATS boards, receive scheduled recommendations, and swipe to manage interest and application status. Runs self-hosted on Cloudflare with private resume storage and no OpenAI API key required.MIT