Weights & Biases

@wandb list all projects under our team entity and show me which ones have been updated in the last 30 days

Found 12 projects under entity 'acme-ml'. Recently active projects include: 'image-classifier-v2' (last run 2 days ago, 847 runs total), 'llm-finetuning' (last run 5 days ago, 203 runs), 'data-pipeline' (last run 12 days ago, 1,450 runs), and 'model-monitoring' (last run 18 days ago, 92 runs). The remaining 8 projects haven't logged runs in over 30 days. This represents the first page of results; pagination may be needed for teams with many projects.

This example demonstrates read-only project discovery. The MCP requires your W&B API key with at least viewer permissions on the entity. Results are paginated (typically 50 per page), so large teams may need follow-up queries. The 'last updated' inference comes from run metadata, not explicit project timestamps.

@wandb show me all artifact types in our 'image-classifier-v2' project and explain what each one tracks

The 'image-classifier-v2' project defines 4 artifact types: 'dataset' (versioned training/validation splits, 23 versions logged), 'model' (serialized PyTorch checkpoints, 15 versions), 'predictions' (inference output CSVs, 8 versions), and 'preprocessor' (feature transformation pipelines, 3 versions). Artifact types organize lineage tracking—datasets feed into model training runs, models generate predictions, and preprocessors ensure consistent feature engineering across environments. This is illustrative; actual version counts depend on your team's logging patterns.

Artifact types reveal your ML pipeline's structure. This read-only query helps audit what's being tracked and identify gaps (e.g., missing experiment configs). The MCP returns type metadata but not artifact contents—you'd need separate queries to inspect specific versions. Useful for onboarding new teammates to existing projects.

@wandb check our team's Slack integrations and draft a summary of which channels are receiving W&B notifications

Your entity 'acme-ml' has 3 Slack integrations configured: Workspace 'Acme Engineering' sends run completion alerts to #ml-experiments (webhook name 'exp-alerts'), model registry updates to #ml-prod (webhook 'registry-prod'), and sweep results to #hyperparameter-tuning (webhook 'sweep-notify'). Based on these channels, your team appears to separate experimental work (#ml-experiments, #hyperparameter-tuning) from production model changes (#ml-prod). Consider whether all three channels need active monitoring or if consolidation would reduce notification fatigue.

This synthesis example pairs W&B integration metadata with AI reasoning about team workflow. The MCP returns channel names and webhook identifiers but excludes sensitive URLs/credentials. Useful for auditing notification sprawl, but note that it won't show you the actual alert frequency or message content—just the configured destinations.

When distributed ML teams need shared training context

A 6-person ML team across two time zones runs long training jobs that span multiple shifts. The morning engineer needs to see what the overnight run produced without digging through Slack or local notebooks. This MCP wins here: the Get Current Viewer and List Projects tools let Switchy pull the team's active experiments and artifact lineage into a shared workspace where the next shift picks up context in under a minute. The List Reports tool surfaces the overnight team's summary without re-running queries. The threshold: if your team isn't already using W&B for experiment tracking, this MCP adds no value—it reads existing W&B state, it doesn't replace your tracking stack. For teams already on W&B who lose 20+ minutes per handoff reconstructing what happened, this is the buy.

When artifact provenance matters more than speed

A 3-person ML ops team at a fintech startup needs to document which models went to production and trace their training lineage for a quarterly audit. They're not debugging experiments—they're answering "where did this artifact come from" under time pressure. This MCP is the right call: List Artifact Types and List Projects give Switchy a read-only view into the registry without touching production infrastructure. The team can query artifact metadata and pull reports into a shared doc that auditors review. The trade-off: this MCP has no write access, so if you need to tag or annotate artifacts during the audit, you're back in the W&B UI. For read-only compliance pulls where the alternative is manual CSV exports, this saves 2-3 hours per audit cycle.

When new hires need W&B context without admin access

A senior ML engineer is onboarding a junior hire who needs to understand the team's experiment structure and automation setup but shouldn't have admin rights to the W&B workspace yet. The MCP's Get Team, List Automations, and List Integrations tools let Switchy surface a read-only map of how the team uses W&B—which Slack channels get notifications, which projects are active, what automations run on new artifacts—without granting the junior full workspace access. This works when onboarding is a 2-day sprint and the senior wants the junior to ask informed questions by day two. The boundary: if your W&B setup is simple enough to explain in a 10-minute screen share, skip this. For teams with 15+ projects and custom automation pipelines, this MCP turns onboarding into self-service.