UploadCheck MCP Server
Thin MCP wrapper over the hosted UploadCheck API. It does not run quality checks locally.
Canonical naming:
- MCP server:
uploadcheck
- CLI/package:
uploadcheck or uploadcheck-mcp
- Tagline: Quality check videos, podcasts, and clips before you upload.
Environment
export UPLOADCHECK_API_BASE_URL="https://api.uploadcheck.app"
export UPLOADCHECK_API_KEY="<workspace_api_key>"
Agent Workflow
Use /check from Claude Code, Codex, or another slash-command capable workspace to quality check a media asset before upload.
/check ./final-upload.mp4
Small and medium local files can be sent through Render without durable storage by base64-encoding the media and passing one of:
video_base64 with video_content_typeaudio_base64 with audio_content_typemedia_base64 with media_content_type and media_kinddata_url
The API writes the payload to a temp file, runs the gate, and deletes the temp file after processing. Use signed URLs or future direct object storage for large files.
The agent should call qc_run_local_file for a reachable local export, or qc_run_video when it already has a YouTube URL, signed URL, upload id, or base64 payload. Then poll qc_get_job, fetch qc_get_report, and list timestamped evidence plus source-level issues it can reach.
Use profile when the caller knows the media type but does not need to hand-maintain every gate. Supported values are auto, nto_long_form, generic_creator_video, shorts, audio, npo_podcast_or_audio, and thumbnail. Explicit checks still win when supplied.
For Codex, Claude Code, Cursor, or NTO/NPO production pipelines, call qc_get_pipeline_handoff first when you need the complete production runbook. It gives the launch preflight, recipe, cost-basis, estimate, media-ingress, report, marker CSV, repair-loop, and rerun sequence. Then call qc_get_pipeline_recipes when you need the current default checks and sidecar arguments, and use qc_run_local_file for the selected profile.
For NPO podcast/audio projects, call qc_get_npo_pipeline_handoff or run uploadcheck npo-pipeline-handoff --json when you need the focused audio handoff without the broader NTO recipe bundle. It includes the cost preflight, default checks, transcript/watchlist/locked-script sidecars, inline/signed media-ingress rules, marker CSV handoff, "Fix now?" prompt, and rerun-before-publish-ready rule:
{
"file_path": "/path/to/final-upload.mp4",
"profile": "nto_long_form",
"manifest_path": "/path/to/storybook.json",
"transcript_path": "/path/to/transcript.txt",
"watchlist_path": "/path/to/watchlist.json",
"expected_script_path": "/path/to/locked-script.txt",
"sidecar_dir": "/path/to/_dialogue-chunks",
"plan_id": "creator",
"cost_guardrail": "downgrade"
}
Small files are base64 encoded by the local MCP process and evaluated through Render inline. Larger files use the signed-upload path automatically unless upload_mode: "inline" or upload_mode: "signed" is specified.
Hosted media ingress can be smoke-tested from the repo before handing the MCP to another production pipeline:
UPLOADCHECK_MEDIA_INGRESS_BASE_URL=https://api.uploadcheck.app UPLOADCHECK_API_KEY=<private_bearer> npm run media-ingress:verify
When a project has a storybook, edit decision list, or visual timeline JSON, pass it as manifest_json with checks: "repeat_fatigue". UploadCheck will use it to flag exact visual reuse and source-family dominance even before a final render is reviewed.
When a project has transcript text or a script-sidecar, pass it as transcript_text or transcript_json with checks: "spoken_leaks". UploadCheck will flag spoken URLs, markdown, prompt text, stage directions, vendor/tool names, and known wrong-name substitutions without running ASR.
For customer-specific terms, pass watchlist_json with checks: "pronunciation_watchlist" plus transcript text. Watchlist entries can include terms: [{ expected, banned: [] }] and top-level banned: [].
When a project has a locked narration script and final transcript, pass expected_script_text or expected_script_json plus transcript text with checks: "script_faithfulness". UploadCheck compares word error rate model-free, so agents can catch narration drift without paying for full multimodal review.
When a project has local chunk QC reports, pass sidecar_dir with checks: "chunk_sidecar_failures". The local MCP process packages JSON sidecars such as *.garble-report.json, Render evaluates them from memory/temp storage, and failed chunk reports become BLOCK flags before upload.
For queued Render worker jobs, do not use inline sidecar payloads because they are deleted after the create request. Use HTTPS sidecar URLs instead: manifest_url, transcript_url, watchlist_url, expected_script_url, and chunk_sidecars_url. UploadCheck fetches them only when the worker drains the job, stores them in temporary server storage for that run, and redacts the URLs from public job/report responses.
For thumbnail candidates, call qc_run_local_file on the image with checks: "thumbnail_text_readability". The same inline Render path evaluates OCR contrast and edge/safe-area readability without model spend.
Use qc_get_cost_basis before pricing, plan, or model-backed review decisions. Use plan_id, ai_review_seconds, and cost_guardrail when an agent is asking for paid AI review beyond deterministic checks. The default guardrail is downgrade: margin-breaking AI review is removed and the job runs deterministic checks. Use block to reject unsafe requests, or off only for internal experiments/deep-review add-ons.
Checked minutes mean deterministic publish-readiness QC minutes. Deep model review stays internal for engine backtesting, roadmap generation, and deterministic capture-rate measurement rather than being sold as a public minute allowance.
Call qc_get_launch_status before launch-sensitive production workflows when an agent needs the live Product Hunt go/no-go state, remaining blockers, and operator commands. Call qc_get_launch_handoff when the agent needs blocker-specific required actions, proof commands, launchDoctorCommands, and the no-launch rule. Call qc_get_launch_doctor when the agent needs the /v1/launch-doctor blocker fix plan plus normalized launch-doctor command coverage, including the hosted Render media-ingress proof command. Call qc_get_launch_evidence when the agent needs a redacted evidence bundle from /v1/launch-evidence without local npm scripts. These launch endpoints are public; the other QC tools still require an API key.
Call qc_estimate_cost before uploading large media or asking for model-backed checks. It returns the effective checks, removed checks, margin safety, and cost estimate without creating a job.
Tools
qc_estimate_costqc_get_launch_statusqc_get_launch_handoffqc_get_launch_doctorqc_get_launch_evidenceqc_get_pipeline_handoffqc_get_npo_pipeline_handoffqc_get_pipeline_recipesqc_get_cost_basisqc_run_videoqc_run_local_fileqc_get_jobqc_get_reportqc_get_eventsqc_get_artifactsqc_get_marker_csvqc_submit_gate_verdictqc_list_recent_jobsqc_create_upload_url
Use idempotency_key with qc_run_video when an agent may retry the same asset. The hosted API returns the existing job instead of creating a duplicate run.
Use qc_submit_gate_verdict after running scripts/qc-engine/run_gate.py so the hosted job report reflects the full-video gate result.
Internal Gemini capture-rate backtests are not part of the public MCP package. Run them only from the private UploadCheck repo with scripts/qc-engine/gemini_watch.py.
Codex Install Shape
cd mcp-server
npm install
node index.mjs
Client Config Snippets
Machine-readable install snippets for Codex, Claude Desktop, Cursor, and generic
stdio MCP clients are published in mcp-install.json. Keep this file aligned
with the package name uploadcheck-mcp, binary uploadcheck-mcp, current
hosted API base URL, and first-call sequence for NPO/NTO production pipelines.
The current MCP install path is a public GitHub clone or local checkout with a workspace API key tied to included plan minutes or an operator-created account. The npx -y @drantoniou/uploadcheck-mcp snippets in mcp-install.json are future-package snippets for after uploadcheck-mcp is published to npm.
