Keep an OpenAI dots Plugin for Lead Finding on Draft Duty

Every lead your dot finds ends in a reply to a stranger, so an OpenAI dots plugin for lead finding stops at the draft and leaves the send with you. OpenAI launched dots on 29 September 2026 with that rule already written into its docs. RedReplier's API has the same rule built in from the other side, which makes the pairing easier to trust than most agent setups.

The short version: connect RedReplier as a custom plugin, give your dot a weekday digest of NEW mentions scored HIGH or VERY_HIGH, let it explain the top three, and keep every status change behind "ask before taking action" in Custom Rules.

Why does lead finding need a stricter setup than monitoring?

Lead finding is the one monitoring job whose output leaves your company, because the reply lands in someone else's thread under your account name. A brand monitoring digest that misreads a thread wastes thirty seconds. A lead finding agent that misreads one puts a sales pitch where nobody asked for it, and the account that posted it wears the result.

The setup itself is the same as for monitoring. The OpenAI dots plugin for Reddit monitoring guide covers developer mode, the https://mcp.redreplier.com/mcp URL, the OAuth sign-in and the read call that proves your dot sees the plugin. Do that first. This guide covers what changes when the output is a conversation with a buyer.

What does "drafts for your approval" mean for a dot?

A dot drafting for your approval means it prepares the words and a person decides whether they go out, and OpenAI states this in three places. The launch post's own example is a dot that "drafts social posts for your approval". The dots docs say "Asking your dot to draft replies doesn't give it permission to send them." The safety post says that "to send a message or share a file, they are taught to seek authorization."

RedReplier holds the same line in its API. Every mention carries an aiReplySuggestion, which the OpenAPI spec describes as a "Draft reply for a human to review before posting", and no RedReplier endpoint or MCP tool posts anything to Reddit, X, Bluesky, Hacker News or Facebook. So the dot reaches a draft through the plugin and has no plugin route past it.

The route that remains is the dot's own browser. A dot browses any site, so it can open a Reddit thread and type. OpenAI's rule means it asks you first. Answer no, every time, and paste the reply yourself after you have read the thread.

What do I ask my dot for each morning?

A lead digest is one list_mentions call filtered to new, high-scoring threads from a fixed window, plus explain_mention on the few worth reading. Save it as a recurring task, which OpenAI's docs describe as a saved schedule.

Every weekday at 08:30, use the RedReplier plugin.
Call list_mentions with statuses NEW, scoreBuckets HIGH and VERY_HIGH,
sort RELEVANCE, and from set to 24 hours ago.
Call explain_mention on the top three.
For each mention give me: source, score, one line on what the person
wants, the link, and the draft reply labelled DRAFT.
Never post, comment or message anyone, on any site.
Never call update_mention_status.
If nothing matches, send one line saying so.

Two details in that prompt earn their place. The from filter works on ingestedAt, the time RedReplier found the thread, not publishedAt, so a thread posted last week and found this morning still lands in today's digest. And "24 hours ago" beats "since the last run", which is the dot's own bookkeeping rather than a field in the API.

Which Custom Rule fits each lead finding action?

Custom Rules decide what your dot does without asking, and lead finding needs a different setting for each kind of action. OpenAI gives a dot four settings: take action without asking, take action when you say so or if pre-approved, ask before taking action, and hand off to you.

OpenAI documents the four settings, not how finely you can scope them, so where your account offers one setting for everything, pick "ask before taking action". A separate auto-review checks planned actions against your instructions, your Custom Rules and OpenAI's safety requirements, but OpenAI has not published how it treats a third-party plugin call as of 29 September 2026. That is why the prompt above repeats the rules in plain words.

Can a dot on ChatGPT Pro mark leads as approved or rejected?

A dot on ChatGPT Pro has no documented route to update_mention_status, because OpenAI's August 2026 help article limits Pro custom MCP connections to read/fetch permissions. The same article says "Full MCP is only available to Business and Enterprise/Edu." It predates dots and uses the older "apps" wording, so test it on your own account before you rely on it.

For lead finding, the Pro limit is the safer default. A rejected mention leaves your NEW queue, so a lead the dot misjudged never comes back and you never learn it happened. On Business Premium the tool is reachable, and "ask before taking action" is the rule that keeps it in your hands.

explain_mention is the one call to test on Pro. RedReplier's spec says it "fills in relevanceReason, tags and aiReplySuggestion when they are missing, which runs AI scoring", which is a write on RedReplier's side. If Pro refuses it, the reason and draft RedReplier already stored still arrive through list_mentions.

What does my dot remember about the people it finds?

Your dot keeps its own memories, including from connected apps, and as of 29 September 2026 you cannot view or edit them. OpenAI's help centre says deleting the dot is the only way to clear them. Every RedReplier mention carries an author field alongside the thread text, so a lead digest feeds usernames and what those people said into that memory.

Two habits keep that small. Ask for the link and the one-line summary rather than the full thread text. And decide before the first run whether you are comfortable with a store you cannot inspect holding a list of strangers who asked for product advice. If you are not, run lead finding from an ordinary ChatGPT conversation with the plugin instead of a dot.

What breaks on lead finding with a dot?

Four things break in the first week, and three of them are about the draft rather than the data.

The first is the draft pasted as-is. An aiReplySuggestion is written before anyone has read the thread's replies, and a thread with twenty answers already needs a different reply than a thread with none. Read the thread, then write.

The second is the bucket read without the reason. Two VERY_HIGH mentions for the same keyword can be a buying question and a two-year-old thread someone revived. explain_mention tells them apart, and our write-up on buyer intent covers what the score measures.

The third is the dot offering to help with the reply in its browser. That is OpenAI's authorization request doing its job. Decline it.

The fourth is the plugin quietly not being used. If the digest arrives without scores, the dot browsed Reddit instead of calling RedReplier. Name the plugin and the tool in the task, as the prompt above does.

How does the dots version compare with Muse and Grok?

The dots version is the only one where the platform itself documents that drafting is not permission to send. The Meta Muse connector for lead finding leans on connector scope and Muse's "Always ask" setting, because Meta does not review custom connectors. The Grok bot template for lead finder runs the same calls from a shared template with a redreplier_ token. All three read the same scored mentions, and the Reddit lead generation page shows what RedReplier does with them before any agent arrives.

Frequently Asked Questions

Can my dot send a Reddit DM to a lead?

RedReplier has no tool that sends a direct message, so the plugin cannot. The dot's own browser can reach Reddit, and OpenAI says dots are taught to seek authorization before sending a message. Set messaging to "hand off to you" in Custom Rules and refuse the request.

Does the digest include a drafted reply?

Every RedReplier mention carries an aiReplySuggestion field, and explain_mention fills it in when it is missing. The spec describes it as a draft for a human to review before posting. Label it DRAFT in the digest so nobody mistakes it for approved wording.

Which platforms can the lead digest cover?

RedReplier's sources filter takes REDDIT_POST, REDDIT_COMMENT, TWITTER, BLUESKY, HACKERNEWS, FACEBOOK and FACEBOOK_GROUP. Leave it unset and the digest covers all seven. Set it to the two Reddit values when you want Reddit alone.

How many leads does one digest return?

list_mentions returns 50 mentions by default and up to 500 with the limit parameter. A lead digest filtered to HIGH and VERY_HIGH that still runs long points to keywords matching too broadly. Tighten the keywords before you tighten the buckets.

Can I use a dots lead finder in the EU or UK?

ChatGPT Pro users in the European Economic Area, the UK and Switzerland cannot use dots as of 29 September 2026. Business Premium includes dots in every supported ChatGPT region. Users under 18 cannot use dots on any plan.

How do I stop the lead finder quickly?

The Activity View shows the dot's current and delegated tasks, and OpenAI's launch post says you can pause or stop the dot's work. To cut RedReplier access, remove the plugin from ChatGPT's Plugins tab. Clearing what the dot remembered takes deleting the dot.

What should the digest do when nothing matches?

The prompt tells the dot to send one line saying so. A one-line answer shows the task ran and found nothing, which is different from a dot that never ran. Keep that instruction in the task text.

Why does a thread posted last week show up in today's digest?

The from filter works on ingestedAt, the time RedReplier found the thread, not publishedAt. A thread posted last week and found this morning counts as new. Use "24 hours ago" rather than "since the last run", which is the dot's own bookkeeping and not a field in the API.

Can I run lead finding in a normal ChatGPT conversation instead of a dot?

You can, with the plugin connected the same way. This is the option to pick if you are not comfortable with a dot's memory, which you cannot view or edit. Deleting the dot is the only way to clear it.

What does it mean when the digest arrives without scores?

The dot browsed Reddit instead of calling RedReplier. Name the plugin and the tool in the task, for example list_mentions with the HIGH and VERY_HIGH buckets. The scores and reasons come from the plugin, not from the browser.

Create your RedReplier account and give your dot a draft-only lead digest tomorrow morning, or read the agent docs first.

Sources: OpenAI's dots launch post, dots safety post, getting started help article, safety FAQ, dots user docs and the August 2026 MCP help article, plus the RedReplier OpenAPI spec and agent docs. Checked 29 September 2026.