Moddit AI

• GIF

  Prompt to create agents on Devvit

• 

  Change how your agent behaves or on what with tools, triggers and instructions

• 

  See your agents behaviour and actions.

• 

  Check Agent instructions

• 

  Test your agent behavior against sample content

• 

  Studio sets example pass and fail cases for Agent context. You can change it or prompt to update it.

• 

  BYOK with any openai compatible inference

• 

  Manage your subreddits agents

Moddit AI turns a few plain-English sentences into working, auditable Reddit moderator agents on top of Devvit, replacing brittle regex with context-aware automation.

About the project

The Gap in Current Moderation Tools Standard moderation tools usually fail because they force a painful trade-off. Traditional AutoModerator regex is incredibly rigid—it cannot read between the lines, look at context, or distinguish a malicious link-drop from an educational resource request. On the flip side, most modern AI solutions are black boxes; they hide their logic behind vague analytics dashboards, forcing moderators to blindly trust an automation they can't inspect. When a tool is either too brittle or too opaque, moderators stop using it, and the backlog piles up in the manual queue.

The Core Insight: Natural Language Intent Moderators aren't prompt engineers, and they don't want to configure complex, nested card settings. They just want to describe the exact behavior they need in plain English and know it will work.

Moddit AI fills this exact gap. A moderator opens the Agent Studio and types their intent naturally—for example: "Catch low-effort founder self-promo dressed up as discussion, but allow honest feedback asks with project context."

How It Actually Works The platform immediately translates that plain text into a structured, executable policy file:

• The Agent Studio instantly builds clear rules, extracts failsafe keywords, and populates explicit "Catch vs. Allow" examples so the moderator can visually confirm what the agent understood.

• The Built-in Simulator allows them to test sample text against the agent before publishing, showing the exact reasoning list and the winning choice ahead of time.

• The Devvit Runtime Integration registers these configurations per subreddit using Devvit's Redis storage (modit:agents:{subreddit}). When a post or comment trigger fires, our decision factory evaluates the content against both deterministic keywords and active LLM inference, ensuring that ambiguous "no-match" results never accidentally override explicit moderation rules.

• The Audit Trail (Agent Actions) streams live operations with expandable, human-readable reasoning reasons so mods always know exactly why a piece of content was flagged, filtered, or approved.

Challenges & Architectural Hardening The hardest part was engineering safety defaults into the trigger runtime. Pure AI inference can be unpredictable, which is unacceptable for live moderation. We heavily leaned into an evaluation factory that balances deterministic failsafe matching with contextual inference. By prioritizing explicit keyword hits and defaulting ambiguous edge cases to a "Review only" state rather than letting them run live, we built an automation framework that actively protects the community while building operator trust.

What's Next We're focused on hardening our end-to-end proofs against live subreddits, adding comprehensive state-cleanup utilities for saved configurations, and expanding our 11+ core template library to handle hyper-specific community engagement behaviors.

Built with

• Devvit (Post & Comment Triggers, Redis Store)

• React & TypeScript

• LLM for Agents

Built With

• devvit
• llm
• react
• redis
• typescript