image-scoring-mcp
Driftara Gallery MCP — one driftara server (search → resolve → execute over local, ipc, cdp, api and system actions); backend work goes to the sibling vexlum (is-be-mcp).
SKILL.md.agent/skills/image-scoring-mcp/SKILL.md
GitHubGallery MCP (driftara)
Catalog: mcp-server/driftara/catalog.json. Contract: DRIFTARA_MCP.md. Quick reference: mcp_tools_reference.md.
Server keys
| Key | Tools |
|---|---|
driftara |
search, resolve, execute; dispatch and sse_status are deprecated aliases |
is-be-mcp (sibling backend, vexlum) |
Same three tools for backend REST, SQL, jobs and diagnostics |
The old gallery keys (is-ui-mcp, is-ui-router, is-ui-local, is-ui-api) still start but warn and serve driftara. Do not use imgscore-el-*, image-scoring-gallery-*, image-scoring-pipeline-*.
Workflows
Gallery startup / backend connection
execute("system.status"): catalog, live server, CDP and backend reachability.execute("local.gallery_status"), thenexecute("api.api_health").- Backend deep dive → vexlum:
search("scoring errors")→execute("diagnostics.get_error_summary", {}).
Anything in the running app (IPC, renderer)
- Start the app with
npm run dev(orENABLE_GALLERY_MCP_SSE=1);ipc.*andlive.*otherwise returnlive_unavailable. search("image count")→resolve("ipc.db.get-image-count")→execute(...).- Renderer:
live.cdp_screenshot,live.cdp_console_logs,live.cdp_click.
Unknown action
search(query) ranks ids, aliases and intent examples. A miss returns suggestions, and a vexlum hint when the intent is backend work.
Writes and destructive actions
writeneedsconfirmed=true.destructiveneeds three things:DRIFTARA_ALLOW_ELEVATED=1in the gallery app's environment;- a
dry_run=truecall that returnsconfirm_token; confirmed=trueplus that token, with the same arguments.
Destructive actions are deletes, import, sync, backup, config save and
live.cdp_evaluate.
Dry-run first. Keep elevation off in automated profiles.