jevjam MCP server

jevjam

Self-hosted MCP server and Jev-compatible HTTP API for small decision models, on one GPU, in Docker.

Ask a model typed questions about a piece of text, JSON, an image or a video, and get calibrated answers back in milliseconds: pick a label (choice), rate on a scale (score), or answer yes or no (noul). Agents call it as MCP tools; services call POST /v1/systemone, the same protocol as TypeSafe Jev, so a Jev client only needs a new base URL.

Use it to route tickets, flag abuse, guard tool calls, pick a model for a prompt, or any other decision you would rather not spend a large LLM call on.

Models

One checkpoint stays in VRAM at a time, and it is freed after five idle minutes. See docs/models.md for sizes, quantization and limits.

Quick start

You need Docker, an NVIDIA GPU, and the NVIDIA Container Toolkit.

docker run -d --name jevjam --gpus all -p 127.0.0.1:8000:8000 \
  -v jevjam-models:/models ghcr.io/beremaran/jevjam:latest

Or, from a clone, docker compose up -d. No model downloads at boot; the first request fetches what it needs into the jevjam-models volume, which takes minutes once.

Ask over HTTP:

curl -s http://127.0.0.1:8000/v1/systemone -H 'Content-Type: application/json' -d '{
  "state": "We were billed twice for March. Refund it today or we cancel.",
  "questions": {
    "department": {"type": "choice", "instructions": "Who should handle this?",
                   "criteria": {"billing": "payments, refunds", "technical": "bugs, outages"}},
    "churn_risk": {"type": "noul", "instructions": "Does the user threaten to leave?"}
  }
}'

The answer, trimmed:

{
  "answers": {
    "department": {"type": "choice", "choice": "billing", "probabilities": {"billing": 0.97, "technical": 0.03}, ...},
    "churn_risk": {"type": "noul", "noul": 0.82, ...}
  },
  "routing": {"model": "english", "reason": "English Latin text", ...}
}

Or connect an agent over MCP, at http://127.0.0.1:8000/mcp:

claude mcp add --transport http jevjam http://127.0.0.1:8000/mcp   # Claude Code
codex mcp add jevjam --url http://127.0.0.1:8000/mcp                 # Codex

Shortened. The full README is on GitHub.