AI Providers (Anthropic, OpenAI, Hugging Face, W&B)

Connect AI provider credentials so Matproof can produce automated EU AI Act and ISO 42001 evidence against your AI training and inference infrastructure.

AI Providers

Matproof’s AI integrations capture evidence about the AI systems your organization develops or deploys — for the EU AI Act, ISO 42001, and the AI-related parts of NIS 2 and DORA. These integrations differ from cloud or identity integrations: they don’t continuously sync data into Matproof. Instead, you store credentials with read-only scopes, and Matproof’s AI Systems inventory module runs targeted checks against your AI infrastructure when you trigger them or on a schedule.

What’s available

Provider	Use case	Evidence produced
Anthropic	Claude API usage tracking	Model versions in production, usage logs, content-filter configuration, system-prompt inventory
OpenAI	GPT API usage tracking	Same as Anthropic — model versions, usage logs, fine-tune lineage
Hugging Face	Open-source model registry, dataset cards, fine-tuning history	Model cards (Article 53 GPAI documentation), dataset provenance, training-data lineage
Weights & Biases	ML experiment tracking	Training runs, hyperparameters, evaluation metrics, model artefacts — for technical-documentation evidence under EU AI Act Annex IV

Why these matter for the EU AI Act

The EU AI Act expects providers and deployers of high-risk AI systems to produce specific technical documentation (Article 11, Annex IV) and post-market monitoring evidence (Article 72). Without these integrations, you’d produce that evidence by hand — exporting screenshots from each provider, pasting CSVs of training runs into spreadsheets, manually tracking model versions through deployment cycles.

Matproof’s AI integrations populate the Foundation Model Cards and AI System Inventory directly from the provider, so your evidence stays current as your models change.

Connecting an AI provider

The connection flow is identical across providers. Each requires only a read-only API key (no OAuth flow — these providers don’t offer OAuth for org-level reads).

1. Generate a read-only API key in the provider's console

   Provider	Where
   Anthropic	console.anthropic.com → Settings → API Keys
   OpenAI	platform.openai.com → API Keys
   Hugging Face	huggingface.co → Settings → Access Tokens (use Read scope)
   Weights & Biases	wandb.ai → User Settings → API keys

2. Restrict the key (if the provider supports it)

   • OpenAI: scope the key to Read-only project permissions
   • Anthropic: scope to a specific workspace
   • Hugging Face: pick Read scope (not Write)
   • W&B: viewer-level access; restrict to specific projects if you only want certain projects scanned
3. Add the credential in Matproof

   Go to Settings → Integrations, find the AI provider, and click Connect. Paste the API key. Matproof tests the connection and stores the credential encrypted.

4. Link to your AI Systems inventory

   Open AI Systems in the sidebar. For each AI system you’ve registered, link the relevant provider credential. Matproof now pulls model metadata, usage logs, and training metadata for that system on the next scheduled scan (or run a manual scan immediately).

What evidence each integration produces

Anthropic

• Models in use — every Claude model version your org has called in the last 90 days, by API key
• Usage logs — request count, token volume, error rate per model (rolled up; no individual prompts retained)
• Content-filter configuration — Anthropic’s safety settings on each workspace
• System-prompt inventory — for AI systems registered with system-prompt tracking enabled

Maps to: EU AI Act Article 13 (transparency), Article 17 (QMS), Article 72 (post-market monitoring); ISO 42001 A.6.

OpenAI

• Models in use — production model versions, including fine-tune lineage (which base model + which fine-tune dataset)
• Usage logs — request volume per project per model
• Moderation flag rate — how often the OpenAI moderation API flagged content over time
• Fine-tune history — datasets used, training runs, evaluation metrics

Maps to: EU AI Act Article 11 (technical docs), Annex IV; ISO 42001 A.6, A.7.

Hugging Face

• Organization model cards — Matproof imports model cards for any model your HF org publishes or maintains. These satisfy EU AI Act Article 53(1)(a–c) GPAI documentation requirements.
• Dataset cards — for training datasets your org owns or uses
• Fine-tune lineage — base model → fine-tune model relationships
• Evaluation results — published evaluation metrics on model cards

Maps to: EU AI Act Article 53 (GPAI obligations), Annex IV; ISO 42001 A.7.

Weights & Biases

• Training runs — every run with hyperparameters, dataset references, evaluation metrics
• Model artefacts — versioned model weights with provenance
• Sweep / experiment results — hyperparameter tuning campaigns
• Reports — W&B reports linked as documentation artefacts

Maps to: EU AI Act Article 11 (technical documentation), Annex IV (records of training/testing); ISO 42001 A.6.2.

Privacy and data handling

These integrations read metadata about your AI usage — not the prompts, completions, or training data themselves.

• Matproof does not store any prompts you send to Anthropic or OpenAI
• Matproof does not download training datasets from Hugging Face (only the dataset card / metadata)
• Matproof does not store model weights from W&B (only the run metadata)

If your AI infrastructure processes personal data, your existing GDPR / DPA obligations apply at the provider level, not at the Matproof level.

Disconnecting

For each provider, Settings → Integrations → [Provider] → Disconnect. The encrypted credential is purged from Matproof’s secrets store.

Also revoke the API key in the provider’s console — disconnection only removes the credential from Matproof; the key itself remains valid in the provider’s system until you delete it there.

EU AI Act

The framework that drives most AI-integration use cases

ISO 42001

AI management system companion standard