Hugging Bay

File-level integrity diff between two artifacts' manifests (added/removed/changed-hash files) — a supply-chain tamper check for re-mirrors or version bumps. A changed sha256 on a same-named file is a red flag. Args: id, against (the other artifact id).

The golden path in one call: recommend open models for a task AND verify each finalist (safe-to-run verdict, effective license, GPU fit, card score, lineage). Returns ranked composite reports plus a bestPick. Use this FIRST when a user asks 'what model should I use for X'. Args: task (e.g. coding, chat, embeddings, vision), commercial (true to require commercial-safe licenses), gpu (e.g. rtx4090-24), limit (1-5).

For a gated or restricted model, return ranked ungated, commercially-usable equivalents at the same capability tier (with hosted-download availability). Arg: id.

Return the single-call agent discovery contract for answer engines, OpenClaw/Hermes-style agents, MCP clients, OpenAPI users, crawlers, citation targets, and bounded API entrypoints.

Return ready-to-use recommendation wording, assistant prompts, integration snippets, citation targets, sharing templates, and claim boundaries for AI assistants and agents.

Return intent-specific agent workflows for RAG, hosted downloads, commercial-safe search, and mirror prioritization.

Return the machine-readable AI crawler/user-agent allowlist, official verification links, and WAF guidance.

Return the AI answer-engine mention monitoring contract, prompt targets, provider targets, citation goals, and crawler readiness signals.

Return public AI mention/citation result rows and summary rollups for answer-engine visibility monitoring.

Return citation-ready AI-search guidance for Perplexity, ChatGPT, Gemini, Grok, Claude, open agents, and MCP clients.

Fetch one answer pack with short answer, canonical citations, extraction APIs, and ranked rows.

Fetch one full artifact JSON document by id.

Fetch one artifact's metadata, card, trust bundle, hosted files, and mirror readiness in one response.

Fetch repo-grade artifact card JSON or Markdown.

Fetch one artifact's hosted download state, upstream source, and reviewed peer-assisted fallback metadata.

Fetch one artifact's fit labels, trust gaps, evidence, hardware fit, and mirror decision data.

Fetch agent-oriented artifact metadata with provenance, hosting, and safe next steps.

Signed, digest-specific artifact passport: provenance, license verdict, safety, bundle state, evidence counts — plus ready-to-paste README embed snippets. Tamper-evident; a changed manifestDigest means the bytes changed.

Fetch public community ratings, reviews, and verified-run attestations for an artifact.

Generate a bay.lock reproducible dependency document for an artifact: immutable hb:// identity, per-file sha256 + sizes + hosting, license verdict, safety snapshot. Restore/verify with the hbay CLI. Arg: id.

Runnable-completeness verdict for one artifact: indexed|metadata-complete|partial-hosted|complete-hosted|run-verified|revoked with the exact missing components. Use before pulling — file counts never imply runnability.

Public bundle-state counts: weightsHosted vs supportFilesOnly vs metadataOnly — hosting reported by runnable completeness, never file counts.

Score a model card's completeness 0-100 (grade A-F) across summary, license clarity, base-model, eval results, datasets, tags, hosted bytes, and provenance, with a prioritized fix list. Use as a card-quality signal or to tell a publisher exactly what to add. Arg: id.

Return public catalog scale, source coverage, and readiness facts.

Typed change timeline for a model: identity/digest changes and advisories, newest first (distinguishes changed vs unsafe). Arg: id.

Chat-template contract for a model: BOS/EOS/stop tokens, thinking markers, tool-calling conventions, and deterministic conformance fixtures — prevents correct weights from producing broken conversations. Arg: id.

Return a compact citation pack for answer engines: safe claims, avoid claims, citation order, proof URLs, and bounded extraction rows.

Moderator-approved community executable stacks (reviewed queue: sha256-pinned dependencies, allowlisted commands; nothing publishes without review).

Fetch aggregate public community trust signals, recent reviews, and top helpful reviews.

Fetch a manifest-backed plan for downloading all Hugging Bay-hosted verified files for one artifact.

Fixed BYOC eval suites for try-before-pull on compute the user owns. Pass suite + endpointKind (openai-compatible|ollama|llama-cpp) + model for an executable request plan — credentials never touch the server.

Fetch public benchmark/evidence rollups grouped by benchmark, tool, and hardware.

Return public health and persistence state.

Dependency blast radius for a model: known catalog derivatives affected if it changes/withdraws + the base-model chain it inherits. Arg: id.

Declared service incidents: open incidents plus incidents resolved within 90 days. Complements the live health/readiness probes.

Fetch sitemap, URL batch, and crawl-readiness status.

Copy-paste-runnable integration snippets for a model across huggingface-cli/huggingface_hub/transformers/ollama/llama.cpp/vLLM/LangChain/LlamaIndex plus a hash-verify recipe and MCP config — filtered to what actually applies to the artifact's type and formats. Arg: id.

Trust-weighted open-model leaderboard for a benchmark (verified evidence outranks self-reported scores); each row carries license clarity. Args: benchmark (e.g. mmlu, humaneval), metric (optional).

Whole-tree license risk for a model: risk grade (low|moderate|elevated|high), effective commercial-use across the base-model chain, laundering signal, and the combined obligations checklist. Arg: id.

Fetch copyable local-use commands, required files, warnings, and next actions for one artifact.

Fetch public manifest signing key metadata.

Fetch one curated pack with artifact rows.

Return mirror candidates with readiness scores, blockers, compatible tools, and selected safe files.

Return a public mirror scaling readiness preview with candidate counts, storage readiness, policy, and blockers. Does not expose execution commands or mirror files.

All normalized benchmark results recorded for one model (benchmark, metric, score, verified flag). Arg: id.

Signed, CycloneDX-compatible AI Bill of Materials for a model: weights files with SHA-256 hashes, license posture with effective commercial-use, base-model dependency chain, and scan/verification state. Archive it or diff it across versions for supply-chain audits. Arg: id.

Model genealogy: exact base-model ancestor chain (with effective-license carry-through) plus known catalog derivatives that name this model as their base. Use to trace provenance or find fine-tunes/quants built on a model. Arg: id.

One-call composite decision document for a model: safe-to-run verdict, license + effective-commercial-use, GPU fit summary, model-card completeness score, and base-model lineage — plus links to every detail endpoint. The fastest way to evaluate a model before recommending or running it. Args: id, gpu (optional, e.g. rtx4090-24).

Standards-backed SBOM for a model: SPDX 2.3 (default), CycloneDX 1.5, or SPDX tag:value — hashed files, license (SPDX id), and base-model relationships, for tools you already run. Args: id, format (spdx|cyclonedx|spdx-tag).

Return popular, newest, and non-Hugging Face open-source AI records in one bounded response.

Fetch one artifact's reviewed torrent/magnet fallback bundle, scoped to hosted-file hashes and reviewer evidence.

Fetch one public publisher namespace profile with releases, hosted files, claim state, and trust stats.

Return a copyable Hugging Bay release manifest template for labs, including source URLs, hashes, runtime notes, upload-session body, and reviewed fallback fields.

List available quantizations for a model (GGUF/AWQ/GPTQ/etc.) with per-quant size, precision, and hosted-download availability, so an agent can pick the smallest that fits. Arg: id.

Fetch one ranking as JSON so agents do not scrape ranking HTML.

Return public mirror signals, totals, and pack summaries.

Fetch a launch-ready release bundle for labs and agents: publisher claim, same-as-upstream evidence, trust bundle, hosted files, local-run commands, and reviewed fallback readiness.

Publisher preflight for a release: objective integrity BLOCKERS (unhashed files, no license, no inventory) separated from documentation-quality WARNINGS (never blocking, never a safety signal). Arg: id.

Fetch a compact release readiness score/checklist for a lab release: namespace, provenance, license, scans, signed manifests, hosted files, runtime guidance, community proof, and reviewed fallback status.

Fetch high-priority artifacts that need community reviews, verified-run evidence, or benchmark reports.

Which local runners can run this exact model (by format + platform) with the run command for each, plus incompatible runners and why. Args: id, platform.

Public security advisories: flagged and quarantined artifacts, open moderation cases, and the scanning policy.

Return one provider's coverage, policy, rollups, and top imported records.

Fetch one stack pack with bounded ranked artifact rows.

Executable open-model stack for a job (laptop-rag, local-coding, multilingual-embeddings, tool-calling-agent, document-extraction): each role resolved to an exact eligible verified variant with lock + run steps. Args: slug, gpu, commercial.

Fetch one topic and ranked artifact rows.

Return privacy-safe aggregate traffic, hosted-download, mirror-request, product-action, and agent/crawler discovery signals.

Return public trending artifacts from aggregate pageview, download, mirror-request, and product-action signals.

Fetch artifact trust evidence.

Evidence missions: ranked verification gaps in the most-used artifacts (first verified run, license pin-down), each with exact submit instructions. Use to direct verification effort where one receipt most improves the evidence base.

Byte-level inspection of a model's hosted files: parses the safetensors header (real param count, dtypes), GGUF header (architecture, context, quant), and config.json (custom-code architecture risk), then cross-checks the bytes against the declared card. Catches mislabeled sizes and card/weights mismatches. Arg: id.

List citation-ready answer packs for AI search engines, agents, and common open AI artifact questions.

List files actually hosted by Hugging Bay.

List one artifact's files with bounded pagination.

Fetch bounded canonical URL batches by kind, with optional artifact source/type/tool filters.

List requested and high-download mirror candidates.

List curated discovery/download packs.

List registered Hugging Bay mirror nodes and health state.

List public publisher namespaces with artifact counts, hosted-file counts, claim state, and trust stats.

List revoked artifact identities (withdrawn/compromised/license-revoked/superseded/malware). Honor these before running or trusting an artifact; a match means do not run.

List practical AI stack packs for high-intent use cases.

List generated catalog topics.

One-call local-run plan for a model: GPU fit verdict, the recommended variant (as-is or the smallest hosted quant that fits), the full quant table, and ordered steps (safety check, hash-verified download plan, ollama/llama.cpp/vLLM run commands, post-download hash verify). Args: id, gpu, ctx.

Migrate off a closed model/API: given a closed model name (gpt-4o, text-embedding-3, whisper, ...), return eligible open alternatives + a migration plan with rollback. Args: from, gpu, commercial.

Preview one hosted text-like file with a bounded first chunk and row samples.

Return ranked artifact recommendations for a concrete use case, with fit reasons, tool commands, and hosted/mirror actions.

Recommend commercially-safe, ungated open models for a use case. Returns a ranked shortlist with license clarity and hosted-download availability. Args: task (e.g. coding, chat, embeddings, vision, reasoning), commercial (true to restrict to commercial-safe licenses), limit.

Paste-ready recommendation PARAGRAPH with inline [n] citation markers resolving to stable Hugging Bay URLs — for when you need to answer 'what model should I use' in prose with checkable sources. Args: task, commercial, gpu.

Resolve an upstream repo URL, owner/name, or artifact id to a Hugging Bay artifact plus hosted/download/mirror actions.

Resolve an immutable hb://{owner}/{name}@sha256:{digest} artifact identity. Returns 'resolved' when current bytes match, or an explicit 'identity-changed' warning — never silently retargets. Arg: uri.

Adversarial-pattern scan of one artifact: unsafe serialization (pickle), payloads hidden among weights, license laundering, homoglyph/typosquat namespaces, metadata spoofing. Returns verdict clean|suspicious|dangerous with per-threat evidence — a verify-before-trust signal, not proof of malice.

Search bounded public catalog rows with summary=1, filters, limit, cursor, and optional semantic reranking.

Will this model run on a given GPU? Returns per-GPU fit verdict (fits/tight/no), weights+KV-cache VRAM estimate, max context that fits, and smaller-quant options. Args: id, gpu (e.g. rtx4090-24, rtx3090-24, a100-80, h100-80, mac-m3-max-64), ctx, batch.