# jevmod

> Open-source content moderation powered by Jev (TypeSafe's System One model). One request per batch of messages, a probability per category (spam, scam, harassment, nsfw, offtopic, selfharm, doxxing, minors) and per plain-language rule, a policy that turns probabilities into flag/delete/timeout. Python SDK, CLI, HTTP API, MCP server, npm package, and Discord/Telegram/Reddit bots over the same core. MIT.

Key facts: flag-only by default; fails open when Jev is unreachable; about $0.000045 per judged message; probabilities move about plus or minus 0.03 between runs; only message text and channel topic are sent to TypeSafe; the key is `TYPESAFE_API_KEY` (keyring, environment or .env, never in code).

## Start here

- [README](../README.md): the three doors (community owner, developer, self-host), commands, categories, the red-team scoreboard.
- [AGENTS.md](../AGENTS.md): exact API signatures, where things live, tests, coding rules, the integration recipe in short form. Read this if you are an agent.
- [BENCHMARK.md](../BENCHMARK.md): jevmod against Llama Guard 3, ShieldGemma and toxic-bert on public datasets, with cost and latency.

## Agent skills and plugin

- [plugin/skills/jevmod-integrate/SKILL.md](../plugin/skills/jevmod-integrate/SKILL.md): procedure to add moderation to an existing codebase: pick the surface, store the key, wire the call, add a real test, decision table, gotchas.
- [plugin/skills/jevmod-moderate/SKILL.md](../plugin/skills/jevmod-moderate/SKILL.md): screen text, inputs and datasets with `jevmod check` (exit codes, --json) or the MCP tools `moderate` and `categories`.
- [plugin/.mcp.json](../plugin/.mcp.json): the MCP server registration (`jevmod mcp`, stdio) that the Claude Code plugin installs.
- [plugin/README.md](../plugin/README.md): what the plugin contains and how to install it from the marketplace.

## Reference

- [jevmod/categories.json](../jevmod/categories.json): the questions and true/false criteria asked to Jev for every category; the single source shared by Python, npm and MCP.
- [jevmod/judge.py](../jevmod/judge.py): the judgment core; its docstring records the red-team findings (dict state, NFKC, character pre-filter, criteria on every question).
- [jevmod/core/policy.py](../jevmod/core/policy.py): default thresholds and actions, `Policy`, `Decision`, `decide`.
- [jevmod/api/server.py](../jevmod/api/server.py): HTTP endpoints and request/response models; OpenAPI at /docs when running.
- [jevmod/mcp_server.py](../jevmod/mcp_server.py): the MCP tools and their signatures.
- [.env.example](../.env.example): every environment variable each role reads.
- [docs/jevmod/](jevmod/): Postman collection requests for the HTTP API, in order of use.
- [docs/diagrams/](diagrams/): architecture, request flow, coverage matrix and failure policy (SVG).

## Project

- [PLAN.md](../PLAN.md): phases, decisions and the production readiness checklist.
- [tests/data/redteam.csv](../tests/data/redteam.csv): 98 labelled adversarial messages used as the regression floor.
- [DISCLAIMER.md](../DISCLAIMER.md): no warranty, no affiliation with TypeSafe, Discord, Telegram or Reddit; the operator owns the actions.
- [LICENSE](../LICENSE): MIT.

## Optional

- [launch/twitter.md](../launch/twitter.md) and [launch/reddit.md](../launch/reddit.md): launch posts with the measured numbers.
- [PRODUCT.md](../PRODUCT.md): the productisation checklist and its owners.