An agent often needs a small judgment before its next action: which workflow fits a request, whether a passage supports a claim, or how relevant an item is under a supplied rubric.
Seb Labs now serves a text-only NVFP4 variant of Seb-9B-v2 through an authenticated API and a remote MCP server. It returns structured judgments rather than generated explanations.
Three ways to decide
A distribution over your labels, for classification and routing.
Σ pᵢ = 1.0000A probability for a precisely defined proposition, for relevance and evidence screening.
P(yes) + P(no) = 1An expected ordinal score and level probabilities under your rubric.
E[s] = Σ sᵢ · pᵢUse these for bounded decisions grounded in the context you supply. Use retrieval for missing facts and code for exact checks. Model probabilities are estimates; they do not replace an uncertainty policy or controls for consequential actions.
Connect your agent
Sign in with Google at the developer console, create a named inference token with an expiry, and use Connect to agent. It copies instructions to install MCP and a persistent decision skill for your agent. Keep the token in the client’s secret environment.
The official agent installation guide supports Claude Code and Codex plugins, plus direct remote MCP setup for other compatible clients:
https://mcp.seblabs.dev/mcp
Authorization: Bearer <your inference token>The free seb_demo checks connectivity using illustrative data. Live decisions require a token and account credits. API and MCP share the same balance and usage history.
What should I use for an exact check?
Use ordinary code for arithmetic, exact validation, and other deterministic checks. Seb is intended for judgments over the context you supply.
Call the API
Send the context, question, and criteria together. This example asks which team should handle a damaged parcel.
curl https://mcp.seblabs.dev/api/v1/systemone \
-H "Authorization: Bearer $SEB_API_KEY" \
-H 'Content-Type: application/json' \
-d '{
"state": "The parcel arrived damaged. Please replace it.",
"questions": {
"team": {
"type": "choice",
"instructions": "Choose the team to handle this request.",
"criteria": {
"returns": "Damaged products and replacements",
"billing": "Payments and invoices",
"escalate": "Unclear or outside these teams"
}
}
}
}'At the time of publication, retail inference pricing was $0.021 per million input tokens, with output free. Each question meters its state, instructions and criteria; failed or invalid inference is not billed. The console shows usage, charges, request timing and errors. At publication, payments used Stripe sandbox and no real-money purchase was required for the preview. See the documentation for current service details.
Vercel AI Gateway status
At publication, Seb Labs had submitted a provider qualification request for native decision-model integration with Vercel AI Gateway. Seb was not yet available in its public catalog, and the upstream contract and provider approval were pending. The integration status page links to the existing API and MCP while that work progresses.
The model card and evaluation notes describe the research evidence and its limits. Hosted NVFP4 service behavior should not be treated as proof of identical performance to the published weight release.
Try a bounded routing or screening task and share the cases where the probabilities are useful, uncertain, or wrong. Documentation · Model card · Console