jevmod

Moderation for communities and apps: every message gets a probability for spam, scam, harassment, nsfw, off-topic, self-harm, doxxing, sexual content involving minors, and for rules you write in plain English. You set the thresholds and the actions. Every decision is logged with its numbers.

It runs on Jev, TypeSafe's System One model: you ask yes/no questions about a message and get probabilities back, no text generation. About $0.04 per 1,000 messages with all categories on.

Site: https://jevmod.dev (community owners) and https://jevmod.dev/developers/ (packages, API, MCP, benchmark, cost calculator)

$ jevmod check "FREE NITRO for the first 100!! claim at discord-gifts.ru/nitro"
scam 0.99      'FREE NITRO for the first 100!! claim at discord-gifts.ru/nitro'  [scam 0.99, spam 0.98, harassment 0.02]

Flag-only by default: nothing is deleted until you turn that on. Fails open: if Jev is unreachable, messages are left alone and the failure is logged. Self-harm is flag-only by design so a moderator can reach out.

you are	you get	start
a community owner, not technical	a Discord bot you tune with commands (Telegram with fewer commands, Reddit by env vars)	Run the bot
a developer	a CLI, a Python package, an npm package, or one HTTP call	Developer
a coding agent, or someone using one	an MCP server and a Claude Code skill that wires jevmod into a codebase	Agents

On the benchmark (2,531 messages from OpenAI's moderation eval, Jigsaw and YouTube spam) jevmod had the best AUROC in every category it was compared on in OpenAI's human-labelled set: harassment 0.93 and sexual 0.98 against Llama Guard 3 8B, ShieldGemma 2B and toxic-bert; self-harm 0.99 and minors 0.98 against Llama Guard, the only other system with those labels. Calibration was measured too; caveats are in the same file. Those four numbers were re-measured on 2026-09-23 after the engine changed under them and they did not move: repeat-run noise on AUROC is 0.001 and the shift was smaller than that. The workings are in the same file.

The key, once

Every surface needs a TypeSafe API key (free tier at console.typesafe.ai).

pip install "jevmod[keys]"
jevmod init          # asks for the key without echo, verifies it with one call, stores it in the OS keyring

Resolution order everywhere (CLI, API, bots, MCP, Moderator): TYPESAFE_API_KEY in the environment, then the keyring, then a .env in the current directory. In a container or a headless server jevmod init --env-file writes .env with mode 600. jevmod init --forget removes the keyring entry (pip uninstall does not). Keys never go into git; .env is ignored.

Run the bot

Discord

1. Developer Portal → New Application → Bot → Reset Token → enable Message Content Intent (the only privileged intent used).
2. OAuth2 → URL Generator → scopes bot + applications.commands; permissions: Read Messages, Send Messages, Manage Messages, Moderate Members, Manage Channels, Embed Links. Open the URL, add it to your server.
3. pip install "jevmod[discord]", set DISCORD_TOKEN, run jevmod discord.

The bot creates a private #jevmod-log channel and starts flagging there.

command (server managers only)	what
/mod status	settings and this month's usage
/mod set <category> <action> [threshold]	any category → off, flag, delete, timeout
/mod rule <name> <text> [action] [threshold]	a rule in your words: "No politics. News about the game is fine." (max 5)
/mod trust <role>	messages from that role are never judged
/mod topic <text>	what this channel is for, in a sentence. Off-topic is judged against it, so it does nothing on its own: turn the category on with /mod set offtopic flag.
/mod log, /mod recent	choose the log channel; last decisions with probabilities
/mod forget, /mod forget_user @member	delete everything stored about the server, or one member

React ❌ on a log entry to mark a false positive (that category's threshold goes up a notch), ✅ to confirm a correct call (down a notch, floor 0.5).

Telegram, Reddit, Twitch and YouTube

Telegram: @BotFather → /newbot, make the bot a group admin, pip install "jevmod[telegram]", set TELEGRAM_TOKEN, run jevmod telegram. Admin commands: /mod_status, /mod_set, /mod_rule, /mod_topic, /mod_log.

Reddit: for your own subreddit with your own "script" app credentials, non-commercial (Reddit's API terms). pip install "jevmod[reddit]", fill the REDDIT_* variables from .env.example, run jevmod reddit. Reports by default; removal and bans are opt-in.

Twitch: register an app at the developer console, then get a user token for the bot account with the scopes chat:read, moderator:manage:banned_users and moderator:manage:chat_messages, and add that account as a moderator of every channel it watches. pip install "jevmod[twitch]", fill the TWITCH_* variables, run jevmod twitch. Chat commands: !jevmod status, !jevmod set, !jevmod rule.

YouTube: OAuth credentials from the Cloud console and a user token with the youtube.force-ssl scope for an account that owns or moderates the live chat. pip install "jevmod[youtube]" (no extra dependency, it is plain REST), fill the YOUTUBE_* variables, run jevmod youtube. It polls one live stream at a time, at the interval YouTube's own response asks for.

Neither of those two can ban anybody, the same as the other three: timeout on Twitch always carries a duration, and there is no code path that omits it. Each adapter's module docstring carries the reasoning.

What the bots send where

Only the message text and the channel topic go to TypeSafe. Author names and ids never do. Locally, jevmod keeps a decision log (category, probabilities, action, first 300 characters); rows older than 30 days are purged on every batch. /mod forget deletes everything; leaving the server does the same. Members whose message is removed get a direct message saying an automated system did it and how to appeal.

Developer

CLI

pip install jevmod
jevmod check "some text"                                   # exit 0 clean, 1 something triggered, 2 error
cat comments.txt | jevmod check --json --rule "No politics. Game news is fine." -

One message per line on stdin, judged in batches of 50. --topic turns on the off-topic check, --threshold sets one for every category, --json prints one object per line with every probability. examples/cli/ has a file screener and a pre-commit hook.

Python

from jevmod import Moderator, Policy

d = Moderator().check("FREE NITRO for the first 100!! claim at discord-gifts.ru/nitro", channel_topic="gaming")
d.action, d.category, d.probability  # ('flag', 'scam', 0.99)
d.scores  # {'spam': 0.98, 'scam': 0.99, 'harassment': 0.02, 'nsfw': 0.01, ...}

p = Policy()
p.set_category("scam", "delete", 0.7)
p.set_rule("no_politics", "No political discussion. Game news is fine.", action="flag", threshold=0.8)
Moderator(policy=p).check_many(["...", "..."], channel_topic="support")  # batched into as few requests as their length allows

npm

import { Moderator, Policy } from "jevmod";
const d = await new Moderator().check("FREE NITRO ...", { channelTopic: "gaming" });
d.action, d.category, d.scores            // same shape as Python and the HTTP API

Node 20+. Same questions (jevmod/categories.json is copied byte for byte and CI fails if it drifts), same policy, same decision. JevmodClient talks to a deployed HTTP API instead, so browsers and edge functions never hold the TypeSafe key. Express middleware in packages/jevmod-js/examples/. Details in packages/jevmod-js/README.md.

HTTP API

JEVMOD_ADMIN_TOKEN=... jevmod api                     # binds 127.0.0.1; JEVMOD_HOST=0.0.0.0 to expose (Docker does)
curl -X POST localhost:8080/v1/keys -H "Authorization: Bearer $JEVMOD_ADMIN_TOKEN" \
     -H "Content-Type: application/json" -d '{"tenant":"my-app"}'          # {"api_key":"jm_...", shown once}
curl -X POST localhost:8080/v1/moderate -H "Authorization: Bearer jm_..." -H "Content-Type: application/json" \
     -d '{"messages":[{"id":"a","text":"FREE NITRO for the first 100!! claim at discord-gifts.ru/nitro"}]}'

endpoint	what
POST /v1/moderate	up to 50 messages → decisions; the X-Request-Id you send comes back as request_id and as a response header
GET/PUT /v1/policy	thresholds, actions, rules for this tenant
GET /v1/decisions	the audit log
DELETE /v1/tenant	forget this tenant
POST /v1/keys (admin)	mint a tenant key, stored hashed
GET /v1/health, GET /metrics	liveness, Prometheus counters

OpenAPI at /docs on a running server (a static copy at docs/openapi.json), a Postman collection at postman/jevmod.postman_collection.json. Any chatbot, forum or comment system that can make an HTTP call can use it; the bots are adapters over the same service.

Examples, one folder per surface

folder	what
examples/sdk/	Moderator basics, a custom policy with a rule, batching 50 per request
examples/cli/	screen_file.sh exits 1 on hits; pre-commit.sh blocks flagged text files
examples/api/	curl, a stdlib Python client, a Node client against a local jevmod api
examples/discord/	run the bot; custom_adapter.py puts any chat platform on ModerationService in 20 lines
examples/agent_harness/	@guarded decorator, Claude Agent SDK PreToolUse/PostToolUse hooks, LangChain callback
examples/input_validation/	FastAPI dependency that answers 422, a pydantic ModeratedText field
packages/jevmod-js/examples/	Express middleware and a plain Node script

Agents

MCP server. pip install "jevmod[mcp]" then jevmod mcp (stdio). Tools: moderate(texts, channel_topic?, rules?) returns one decision per text; categories() describes each category and its default threshold.

claude mcp add jevmod -- jevmod mcp                                   # Claude Code
{"mcpServers": {"jevmod": {"command": "jevmod", "args": ["mcp"]}}}    # Cursor, Codex, others

Claude Code plugin. This repository is its own marketplace:

/plugin marketplace add ohernandezdev/jevmod
/plugin install jevmod@jevmod

jevmod-integrate adds moderation to an existing codebase (detects the stack, picks SDK/npm/HTTP/MCP, wires the key, inserts the call where it belongs, adds a real test). jevmod-moderate screens text or datasets from the terminal while working. AGENTS.md has every signature; docs/llms.txt indexes the docs.

Categories

category	true when	default
spam	unsolicited promotion, invite farming, bare link drops, mass mentions	flag ≥ 0.85
scam	fake giveaways, phishing domains, impersonated support, "DM me for a deal"	flag ≥ 0.75
harassment	insults, slurs, threats, targeted abuse, in any language	flag ≥ 0.75
nsfw	sexual or gore content for a general audience (below the threshold means SFW)	flag ≥ 0.80
offtopic	unrelated to channel_topic; needs a topic to mean anything	off, 0.90
selfharm	the author is in crisis or considering self-harm; alert moderators, never punish	flag ≥ 0.50
doxxing	reveals or hunts private data about a real person	flag ≥ 0.80
minors	sexualises a minor or shows grooming behaviour	flag ≥ 0.70
rule:<name>	your rule in plain language, exceptions included, up to 5	flag ≥ 0.80

Every check returns all enabled categories at once, in one request. The questions are in jevmod/categories.json: one yes/no question per category with explicit true/false criteria, the pattern of TypeSafe's guardrails cookbook. Jev's probabilities move about ±0.03 between runs, so anything within that band of a threshold will flip; the ❌/✅ feedback and PUT /v1/policy exist to move the line.

How it works

Message → pre-filter (trusted authors, under eight letters without a link, repeats of judged text never reach Jev) → batch for 2 s per community → one Jev request for the batch → probabilities → policy → action → audit log. Batched messages are sent as a dict keyed by position; as a list, probabilities leaked between neighbours, which the 98-message adversarial red team caught and which now runs as a regression suite in CI. More in docs/diagrams/ and PLAN.md.

Self-host

One image, one variable picks the role: api, discord, telegram, reddit, twitch, youtube. SQLite on a volume.

cp .env.example .env && docker compose up -d              # API on :8080
docker compose --profile discord up -d                    # add the Discord bot

A $4/month VM, Fly.io or Railway with a volume is enough. Failure policy: Jev unreachable → decisions come back reason="error_open" and nothing is acted on. There is no quota by default; JEVMOD_MONTHLY_QUOTA=5000 pauses judging for a tenant after 5,000 judged messages in a month as a cost guard, tells the owner once, deletes nothing.

Public demo

The landing page's live check talks to jevmod demo, a separate role that keeps the key on the server and stops at a monthly budget (JEVMOD_DEMO_BUDGET_USD, default $0.50), with per-visitor limits and a CORS allow-list. It logs what visitors try (text, scores, hashed IP) for the operator. deploy/demo/ has a Caddy + Docker compose for a small VPS with HTTPS in two commands.

Development

git clone https://github.com/ohernandezdev/jevmod && cd jevmod
python -m venv .venv && .venv/bin/pip install -e ".[all,dev,examples]"      # Windows: .venv\Scripts\pip
ruff check . && mypy jevmod && pytest          # offline tests run without a key; the rest hit the real API
cd packages/jevmod-js && npm ci && npm test    # same for the npm package

Tests never mock Jev. tests/test_redteam.py is the adversarial regression set; benchmark/ reproduces BENCHMARK.md.

License

MIT, © Omar Hernandez. See DISCLAIMER.md: decisions are probabilistic, the operator owns the thresholds, the actions and legal compliance; not affiliated with TypeSafe, Discord, Telegram or Reddit.