OpenJEV

What OpenJEV MCP server does

The OpenJEV MCP server gives an MCP client structured judgment tools for decisions that an agent or application needs to act on. It sends a state value—such as text, an object, or an array—along with instructions to the OpenJEV API, which is powered by Jev, TypeSafe's System One model.

Results are designed for downstream logic rather than conversational output. Depending on the tool, a response can identify one option, place an item on an ordered scale, express a yes/no probability, or return several independent answers keyed by identifiers. This makes the server suitable for support routing, content moderation, change-risk review, lead qualification, evaluation, and similar classification tasks.

How it works

An MCP client starts the server as a local stdio process. The API key remains in that process's environment instead of being passed through the model, client, or tool arguments. Each request is checked locally before transmission. The server can retry temporary failures and validates the returned judgment before exposing it to the client.

A single request can contain one context and multiple independent questions. For larger collections, the batch tool applies the requested judgments to separate items and reports item-specific errors. The API origin, model alias, timeout, retry count, and batch concurrency can be adjusted with environment variables.

Setup and configuration

OpenJEV MCP server requires Node.js 20 or newer and an API key from the OpenJEV dashboard. The usual configuration runs npx -y openjev-mcp and supplies OPENJEV_API_KEY in the server environment. Claude Code, Cursor, VS Code, Claude Desktop, Windsurf, and Cline are documented configuration options.

A key can be tested without starting an MCP session by running npx -y openjev-mcp --check with the key set. A valid check exits successfully after making one judgment; a rejected key exits with an error, and a missing key uses a separate failure code. The process writes JSON-RPC only to stdout, with startup errors sent to stderr.

For pinned installations, install the package globally and run openjev-mcp, or build from a checkout and invoke the compiled JavaScript file with Node. If using a .env file, pass Node's --env-file option before the compiled script rather than as an argument to the package command.

Tools and capabilities

• jev_ask answers several independent questions for one shared context. Its result is an answers object keyed by question IDs.
• jev_choice selects from caller-defined options and returns the selected option, probabilities, and confidence.
• jev_score evaluates an ordered scale and returns a score, legend, probabilities, and confidence.
• jev_noul returns a value from 0 to 1 for a yes/no judgment.
• jev_batch applies questions across multiple records, returning answers keyed by item IDs and per-item errors.

Every tool accepts instructions and an optional model. The batch tool also accepts a concurrency setting, allowing callers to limit the number of requests kept in flight.

Limitations and notes

The service needs a valid OpenJEV API key and network access to the configured API origin. Calls can take long enough that clients with a short MCP timeout may stop waiting; the documented DeepSeek Harness setup raises its timeout because the server can make multiple attempts. The server exits immediately when the key is missing or a setting cannot be parsed.

The tools produce typed judgments and probabilities, not general-purpose prose responses. Select jev_choice for a defined set of alternatives, jev_score for an ordered scale, and jev_noul when the probability of a yes/no condition is the required value.