Purpose
jev-mcp-server is the MCP server for Jev. Jev is TypeSafe's System One model, a hosted decision model. It receives structured questions with enumerated options and returns typed answers with calibrated probabilities and confidence values. It does not write prose text, nor does it explain itself.
This server is used to connect Jev to MCP-capable clients or agents, enabling them to call Jev for small, repetitive judgments, such as triage, routing, scoring, and gating; these tasks are often too slow and too expensive if every item is handled by a frontier model.
Main capabilities
The MCP server exposes Jev's three official question types:
- choice: choose among the given options.
- score: score according to a rubric.
- noul: one of the official question types (the original text does not expand its definition; no further explanation is currently provided).
It also provides:
- compare: A/B comparison.
- verify: checking assertions against evidence.
- batch classify: batch classification.
- an installer that writes the client configuration with a single command.
Examples given in the official text include:
- choice: which change is most likely to cause a breaking change for API consumers? Options: rename an existing configuration key, add an optional response field, change the log format.
- score: the regression risk of rewriting the authentication middleware in place without writing tests, using a mild/moderate/severe scoring rubric.
- verify: assertion "all tests in the latest CI run passed", evidence "the CI log shows 3 of 250 tests failed".
Typical applicable tasks (according to the official text):
- enumerated options: triage, routing, tie-breaking, A/B calls
- rubric scoring: risk, severity, review triage
- high-volume binary checks: breaking change gating, assertion and evidence checking
- batch labeling: 100 items against one label set, about $0.007
Usage
A TypeSafe API key is required before use. There is no local model, no second LLM, and nothing else needs to be installed.
Installation preparation: if uv is not yet installed, you can run curl -LsSf https://astral.sh/uv/install.sh | sh, or brew install uv, or pip install uv. You can also directly pip install jev-mcp-server; in both cases the installer subcommand is the same.
Most stdio MCP clients use a similar form. Codex (~/.codex/config.toml): the official text does not give specific configuration content; none is currently provided. OpenCode (opencode.json the "mcp" section):
{"jev": {"type": "local", "command": ["uvx", "jev-mcp-server"], "env": {"TYPESAFE_API_KEY": "apikey_xxx"}}}
If you do not want to write the key into the configuration, you can skip the env block and call the setup tool once from the agent. It validates the key in real time and stores it with 0600 permissions.
Cost and latency
The official text states: a single call costs about $0.00002 to $0.0001 and returns in about 0.5 seconds; cost is calculated at $42 per 1 billion input tokens (officially marked as indicative). After setting JEVMCP_CACHE=1 , the same question payload will be answered from disk at zero cost (usage.cached: true); classify automatically deduplicates repeated entries.
Applicable scenarios
- When you need fast, cheap duplicate determination for a small number of entries.
- When you need to do triage, routing, scoring, or gating in an agent loop.
- When you need batch labeling, for example 100 entries against one label set.
- When you need to do binary verification of assertions against evidence.
Limitations
According to the official text, the following cases are not applicable:
- Open-ended reasoning or long-context analysis.
- Any scenario that requires accompanying explanations. Jev returns numbers; any "reason" attached by the assistant is its own interpretation of those numbers, not the model's output.
- Jev never explains its choices, and this is by design; "decisions, not strings" is the product positioning, and the probability distribution is the output. Explanations would consume the latency and tokens that this model is meant to avoid.
Other notes: Jev is a cloud API, with no local model and no auxiliary LLM; it is up to the agent's main model to decide when to call these tools and read the numbers. If it is just a one-off question, asking the main LLM directly is also fine; the difference shows up in loops - the LLM's self-reported confidence is uncalibrated, a single call is about 100 times more expensive, and it takes seconds rather than milliseconds. In addition, there are other Jev MCPs: jkudish/jev-mcp (Node/npm) provides ten workflow-shaped tools, and there is also a Go server; this project is on the Python/uvx side, faithfully implements the existing official question types, installs with one command, has bilingual documentation, and is MIT licensed.
