Skip to main content
Glama
jev-domain-mcp

jev-domain-mcp

by jiawei686
README.md
# jev-domain-mcp

> ⚠️ **Superseded by single-purpose servers.** This aggregated "two tools in one
> server" repo is kept only as a *reference example* of the bundled approach.
> The recommended layout is **one MCP = one function**. Use these instead:
> - [`jev-review-mcp`](https://github.com/jiawei686/jev-review-mcp) — code-review gate (`review_patch`)
> - [`jev-screen-mcp`](https://github.com/jiawei686/jev-screen-mcp) — content-moderation gate (`screen_content`)

Domain-specific MCP tools powered by [TypeSafe Jev](https://typesafe.ai) — the
System One decision model. The agent gets **typed decisions, not prose**:
each tool pre-builds the Jev questions (candidate sets, rubric, thresholds), so
the host never needs to know Jev's API.

This fills the gap identified in the Jev ecosystem: the *generic* Jev MCP layer
is already crowded (`burnigtm/jev-mcp`, `hangarbay/jev.mcp`, `mcp-typesafe`),
but the *domain* apps in the gallery are standalone CLIs — none expose their
decision function as an MCP tool. This server wraps two of them.

## Tools

| Tool | Domain | What it decides |
|------|--------|----------------|
| `review_patch` | Dev workflow | Patch verdict (approve / request_changes / needs_discussion), safety probability, severity, **auto_merge vs human_review** |
| `screen_content` | Content safety | Spam & toxicity probabilities, category, severity, **allow / review / block** |

Both return a `confidence` (calibration concentration, **not** truth) and an
`action` the agent can branch on directly.

## Install

```bash
npm install
npm run build
```

Set your key (optional — without it the server runs in deterministic **mock**
mode, no network):

```bash
export TYPESAFE_API_KEY=ts_xxx      # live Jev calls
# or
export JEV_MCP_MOCK=1               # force deterministic mock (no key needed)
```

Diagnostics:

```bash
node dist/index.js doctor
node dist/index.js doctor --json
```

## Use

Register the server with any MCP client (Cursor, Codex, Claude Code, WorkBuddy):

```json
{
  "mcpServers": {
    "jev-domain": {
      "command": "node",
      "args": ["/absolute/path/to/jev-domain-mcp/dist/index.js"],
      "env": { "TYPESAFE_API_KEY": "ts_xxx" }
    }
  }
}
```

Then the agent can call `review_patch` and `screen_content` like any other tool.

## How a tool works

```
state (diff / text) ──► Jev: one request, parallel typed questions ──► structured result
```

- `Noul` → 0..1 probability (is it safe? is it spam?)
- `Choice` → pick from a fixed candidate set (verdict, category)
- `Score` → ordinal rubric (severity)

The result includes `probabilities` + `confidence`, and an `action` derived
from a simple confidence gate. **Low confidence always routes to a human /
stronger model** — that is the whole point of a calibrated decision model.

## Constraints (from Jev)

- Pure text input: no image / audio / video.
- State + all questions must fit ~64k tokens total; a single question ≤ 32k.
- Do arithmetic / date math in host code, not in Jev.
- Output is **free**; you pay only for input tokens.

## License

MIT

TDQS

A4.1/5.0

Scored across 2 tools

Disambiguation5/5

review_patch and screen_content operate on entirely different inputs (git diffs vs. text) and produce different domain-specific verdicts. There is no realistic ambiguity about which tool to select for a given task.

Naming Consistency5/5

Both tool names follow the same lowercase verb_noun pattern: review_patch and screen_content. The naming is concise, predictable, and clearly conveys the action being performed.

Tool Count3/5

Two tools is at the low end of acceptable for an MCP server. Each tool is substantive, but the server feels thin for something labeled a 'domain' MCP and would benefit from additional related decision gates.

Completeness3/5

The two exposed gates are self-contained for their specific inputs, but the broader Jev decision-model domain is not fully mapped. There are no tools for configuration, audit feedback, or other content types, which could be notable gaps depending on the intended use.