Regulars: Turn your chat into a crew.

We developed Regulars using OpenAI Codex as our coding partner throughout the project. It is a local Twitch companion that helps streamers follow conversations, welcome returning viewers, and turn shared moments into community traditions.

The inspiration came from a simple problem: a returning viewer, a thoughtful question, and the beginning of an inside joke can all disappear in the scroll while a streamer is busy playing. We wanted to give streamers useful context at the moment they need it, helping them build relationships while keeping up with the stream.

Regulars brings five connected features into one private dashboard and OBS overlay:

Lurker Bridge provides private reminders about regulars, returning viewers, and first messages, helping the streamer find a natural moment to say hello.
Chat Organizer groups messages into conversations, making overlapping discussions easier to follow.
Chat Stickers turns selected conversations into generated stickers. The streamer reviews and approves each sticker before viewers can trigger it on the overlay.
End-of-Stream Awards recognizes discussion starters, late-night participants, and early supporters.
Guardrails lets the streamer describe unwanted topics and choose whether matching messages should disappear from their dashboard or be removed from Twitch chat.

These features share the same community memory. A conversation discovered by the organizer can become a sticker, contribute to an award, and provide context for a future reminder.

We built the backend with Node.js and SQLite, and the dashboard and overlay with JavaScript, HTML, and CSS. The Twitch adapter implements OAuth, EventSub WebSockets, and Helix API calls. A Python OBS hotkey adapter supports navigating the reading interface while another application has focus.

We divided the AI work by purpose. Jev handles classification and relevance judgments. GPT-6 Luna names conversations, summarizes chat, and creates sticker briefs. An image model generates the sticker artwork. Application code manages attendance, timestamps, counts, and award eligibility.

Codex helped us turn the product requirements into working implementations across that stack. We used it to develop backend modules, connect the interfaces, refine the narrow OBS dock, debug topic grouping, and build recording and replay tools. As failures surfaced, we worked with Codex to investigate them and add regression tests. The development cycle became: describe the intended behavior, implement it with Codex, replay a conversation, inspect the result, and refine the implementation.

The hardest challenge was keeping conversations distinct as chat evolved. Two discussions can share vocabulary while covering different subjects. A short message such as “How much did it cost?” depends on its reply context. Topics also emerge gradually, so earlier messages sometimes need reconsideration once more evidence arrives. We addressed this through reply context, competing topic scores, confidence thresholds, and short subject labels for messages that remain ungrouped.

Another challenge was coordinating asynchronous work. Messages keep arriving while models respond, and a message can be deleted while processing is underway. We added bounded queues, retries, cancellation, and checks against stale results. Privacy influenced the architecture because a removal needs to propagate through the context and derived records that used that message.

We learned that reproducible examples make both AI development and collaboration with Codex more effective. Recording and replaying chat let us investigate quiet conversations, sudden bursts, overlapping subjects, and delayed responses under repeatable conditions. We also learned how much a streaming tool depends on readable cards, predictable shortcuts, and clear approval steps.

Regulars is currently a working local prototype with a replayable demo. Our next milestone is validating the Twitch integration and OBS workflow on a live channel, then using streamer feedback to improve how it supports real communities.