MCP
Six remote tools, not a chat model
This MCP server is a remote tool socket, not a chat model. After a client connects, the agent sees six jev_* tools and calls one at a decision boundary. The fields match that site's REST API. The server is www.jevai.org/api/mcp, and the write-up is the community MCP page. JEV Hub does not host this server.
- One call at the boundary
- POST JSON-RPC on HTTPS www
- A Bearer key, not a login cookie
- The six tools
- A probability is not permission
One call at the boundary
The agent keeps its own model. At a decision boundary it calls one of the six tools. The reply is a structured judgment, not a paragraph, and not a tool the server runs for you.
Field names line up with the REST routes below. This hub does not add another tool and does not proxy the calls.
POST JSON-RPC on HTTPS www
The endpoint is stateless Streamable HTTP. The client sends POST JSON-RPC. GET is not an SSE stream.
Use HTTPS and the www host: https://www.jevai.org/api/mcp. Plain http and the bare domain redirect. Many clients then resend that POST as a GET and receive 405.
A Bearer key, not a login cookie
Authentication is a personal Bearer key. A login cookie is not enough. initialize, ping, and tools/list do not call Jev and do not use the daily quota. tools/call does.
The protocol version is 2025-03-26. The server name is jevai. Keys still come from jevai.org/agent/keys, not from JEV Hub.
The six tools
Call one tool when its boundary matches the moment. A returned probability is a signal, not an authorization.
jev_guard_tool_call
Before a consequential tool call. Returns allow, confirm, review, or deny. It does not run that tool. Required: tool and action. Optional: arguments_summary, side_effects, safeguards, policy, and reversibility. REST: POST /api/v1/decisions/tool-guard.
jev_route_model
When several models can do the work, and quality, cost, latency, or tool access would change the pick. Required: task and candidates. Optional: priorities, constraints, and stakes. REST: /decisions/model-route.
jev_route_task
When the next path is unclear or risky. Returns proceed_fast, deep_review, split_task, or block. Required: task. Optional: evidence and constraints. REST: /decisions/route.
jev_check_research
The evidence is already in hand. Judge one claim as accept, verify_more, or reject. It does not browse. Required: claim. Optional: evidence, source_quality, and stakes. REST: /decisions/research.
jev_review_completion
Before declaring a non-trivial goal finished. Returns complete, verify_more, or incomplete. Required: objective. Optional: completed_work, verification, and known_gaps. REST: /decisions/completion.
jev_decide
When no preset fits, send state and choice, noul, or score questions. Required: state and questions. Optional: model, and that value can only be a Jev model id. REST: POST /api/v1/decisions. The question types are Choice, Score, and Noul on this hub.
A probability is not permission
The numbers in the reply are a signal, not an authorization. They do not skip an approval the user or the platform already required.
This hub does not host this server. See how an agent asks for a judgment, the Skill files that name these tools, and the request builder that does not call an API.