Capabilities · Recommendations
Product Recommendations
HealthyIn-context predictionEndpoint product-recommendationsProduction v12Challenger v13
Requests / day
782K
CTR
8.7%
Conversion
3.2%
Revenue influenced
$412K
Production
2 weeks ago
v12
- NDCG@10
- .412
- P95
- 58ms
- Cost / 1K
- $0.22
Serving 80% of traffic · A/B running
Challenger
4 days ago
v13
- NDCG@10
- .421
- P95
- 61ms
- Cost / 1K
- $0.23
Shadow evaluation complete · A/B: 20% traffic
Experiment
Running
Call it from your product
Sandbox
items = ai.recommend(
capability="product-recommendations",
user_id="usr_1282",
limit=10
)Response
{
"prediction_id": "pred_01J9X4R3TC",
"items": [
{"item_id": "sku_44120", "score": 0.83},
{"item_id": "sku_10982", "score": 0.79},
{"item_id": "sku_77013", "score": 0.74}
],
"version": "v12"
}Next step
v13 beats v12 on NDCG@10 (.412 → .421) in offline evaluation. Shadow results confirm it on live traffic.
Activity
Today
- 07:30Experiment v12 vs v13 — day 4 of 14, 20% challenger traffic
4 days ago
- 13:05v13 created — session events added to context
Last week
- 09:12Managed B rejected — P95 112ms violates constraint
2 weeks ago
- 15:40v12 promoted to production
Use it from an agent
Every capability is also a tool. Give your agents calibrated predictions instead of letting them guess.
Private preview
One remote MCP server for the whole workspace. Works with Claude, Cursor, VS Code and any MCP client.
Claude Code
claude mcp add --transport http zerograd https://mcp.zerograd.dev/v1 \
--header "Authorization: Bearer $ZEROGRAD_API_KEY"mcp.json · Claude Desktop, Cursor, VS Code
{
"mcpServers": {
"zerograd": {
"type": "http",
"url": "https://mcp.zerograd.dev/v1",
"headers": { "Authorization": "Bearer ${ZEROGRAD_API_KEY}" }
}
}
}Tools exposed
- product_recommendations_recommendRecommendationsReturn a ranked list of products a user is likely to buy.
- get_capabilityHealth, production version and live quality metrics.
- list_capabilitiesEvery capability in the workspace.
- record_outcomeSend the real outcome for a prediction_id once it is known.
Tools are scoped by API key. Calls are logged, metered and show up in this capability's observability like any other request.