Build One Meta Muse Connector for Social Listening, Not Four

Reddit, Hacker News, X and Bluesky have one thing in common inside Meta's agent, which is why a Meta Muse connector for social listening is something you build rather than something you install. Muse ships a connector for none of them. As of 20 September 2026 none appears in any connector list Meta or its executives have published, and one is owned by a direct Meta competitor.

The short version: build one Custom Connector against RedReplier's public OpenAPI spec at https://redreplier.com/openapi.json and all four sources arrive in a single GET /mentions call, already scored. Four separate scrapes become one bearer token.

Why does one connector beat four?

One connector beats four because the four sources are already merged before Muse sees them. A Custom Connector is the route Meta documents for a service not yet available in the Connector list. Parallel watched one get built on 14 September 2026: Muse read the documentation, wrote the client, tested the calls and saved the integration as a reusable skill. The Reddit monitoring walkthrough covers that mechanism end to end, including where the credential lands and what Meta reviews, which is nothing.

Four scrapes means four auth flows and four things that break independently. RedReplier's GET /mentions returns all four in one response shape, plus Facebook, which is the one source Muse already reaches on its own. Filter sources to the four Muse cannot see and "what did people say about us today" and "what did people say on Bluesky today" are the same call with a different value.

Which of the four sources can Muse reach by itself?

None of them, as of 20 September 2026. The full connector roster and where it comes from is in the pillar. The list runs to email, calendar, banking, restaurants, music, health and fitness, and it contains no social listening source at all.

X deserves its own sentence, and the X-only walkthrough gives it one: Meta and the owner of X compete directly, so the absence is structural rather than a backlog item. Hacker News and Bluesky are the opposite case, small enough that nobody has built for them, which lands you in the same place.

Two fallbacks remain, and both are worse. The pillar covers why the browser route and the missing MCP field both dead-end. Four sources multiply the first one: browser automation that breaks on a layout change breaks four times as often across four layouts.

What goes in the setup message?

The setup message names the spec URL, the auth style, the scope and the first filter you want. RedReplier's OpenAPI 3.0.3 document needs no auth to read, which keeps this to one message.

The pillar has the full setup message, spec URL and bearer header included. One line changes here:

Use GET endpoints only for now. First call: GET /mentions with
statuses=NEW and sort=RELEVANCE, no sources filter, so all four
arrive together. GET /mentions/count returns a single total, so call
it once per source and show me the four counts before any text.

Create the token at redreplier.com/api-tokens, where it comes back prefixed with redreplier_. It belongs in the secure credential prompt, because a token typed into the chat stays in the chat.

How do I stop the agent reading four hundred mentions?

You stop it with relevanceScore, the 0 to 100 number RedReplier attaches to a mention once the scorer has read it. Four sources produce four times the volume of one, and an agent handed the raw feed summarises the noise as carefully as the buyers.

Pass sort=RELEVANCE rather than sort=RECENT. Recency is where multi-source monitoring fails: Reddit and X produce far more matches per hour than Hacker News or Bluesky, so a feed sorted by time buries the slower sources. Relevance flattens that, and a Hacker News thread scored 90 arrives above an X post scored 30 whatever time either landed. The buyer intent score write-up covers what the number is measuring.

What does the explain call add that the score does not?

POST /mentions/{id}/explain returns the reasoning behind the number, which is the difference between a score you trust and a score you argue with. A 72 on a Hacker News comment and a 72 on a Reddit post are the same number for different reasons.

It is the one POST worth naming in a GET-only setup message, for the reason the pillar sets out. Across four sources the payoff is larger: the reasoning is what makes the two 72s comparable. Everything else that writes stays out: PATCH /mentions/{id}/status rejects mentions that then leave the queue, and POST /keywords/activate-pending charges your account.

How often does each plan check the four sources?

Cadence is a plan setting, and GET /alert-settings returns cadenceMinutes alongside minIntervalMinutes, the floor your plan allows. The pillar walkthrough covers what each plan costs. What matters across four sources is the floor: 60 minutes on Starter, 15 on Business.

A Reddit thread four hours old has been answered by the time you open it, while a Hacker News thread of the same age is still on the front page. Read minIntervalMinutes from GET /alert-settings and match it to your fastest source, not your keyword count. PUT /alert-settings rewrites it, so keep that call out of the connector.

What breaks across four sources specifically?

Five things break, and four only show up once more than one source runs through the same agent.

A null relevanceScore is the first, and an agent sorting numerically on one either drops that mention or parks it at an end of the list.

Source skew is the second. Reddit and X dominate raw volume, so a digest capped at ten items runs all Reddit for a week while the Bluesky thread you needed sits at position eleven. GET /mentions/count returns one total for whatever filter you pass, so call it once per sources value first and you see the skew instead of inheriting it.

The reply rule is the third, and it holds on all four. RedReplier drafts the reply and a person sends it, because unreviewed replies get accounts banned, and Reddit is the least forgiving of the four about it.

Drift between the saved skill and the API is the fourth. Parallel found on 14 September 2026 that Muse saves a custom connector as a reusable skill that persists across conversations, holding whatever endpoints it learned the day it was built. No published source documents what Muse does when that API changes, so re-point it at the spec after any change.

The fifth is the VM's location, which the pillar covers. What it rules out here specifically is bridging: the Grok bot route to the same four sources runs on your machine, so the two cannot feed each other.

Frequently Asked Questions

Does one connector really cover all four sources?

Yes. GET /mentions returns Reddit, Hacker News, X and Bluesky in one response, and the source is a field on each mention rather than a separate endpoint. Filtering to one source is a parameter change, so a social listening setup that starts cross-platform narrows later with no rebuild.

Does a null relevanceScore break the digest?

A null relevanceScore breaks numeric sorting, not the call. The spec marks relevanceScore nullable, so a mention fetched before the scorer reaches it arrives without a number. Tell the connector to return unscored mentions as their own group rather than sorting them to an end.

Can Muse post a reply to X or Reddit through this connector?

No. RedReplier's public API exposes no endpoint that publishes a reply, on any of the four sources. The connector drafts and surfaces, and you send.

Why is there no Muse connector for X?

As of 20 September 2026 X appears in no Muse connector list Meta or its executives have published, and Meta has given no reason for the absence. The X-only walkthrough covers what that means for a connector built on sources=TWITTER. RedReplier's GET /mentions already returns TWITTER rows, so one connector reaches X without an X API tier of your own.

Which source floods the digest first?

Reddit arrives under two source values, REDDIT_POST and REDDIT_COMMENT, while Hacker News and Bluesky each have one, so a cap written per source has to cover both or half the Reddit rows slip past it. Call GET /mentions/count per source first and cap per source rather than per digest.

Where do I get the token for the connector?

Create it at redreplier.com/api-tokens, where it comes back prefixed with redreplier_. Paste it into the secure credential prompt, because a token typed into the chat stays in the chat.

Should I sort mentions by recency or relevance?

Sort by relevance with sort=RELEVANCE. Reddit and X produce far more matches per hour than Hacker News or Bluesky, so a feed sorted by time buries the slower sources. Relevance puts a Hacker News thread scored 90 above an X post scored 30 whenever either landed.

Does Meta review a custom connector before I can use it?

Meta does not review custom connectors or how they use your information. A Custom Connector is the route Meta documents for a service not yet in the Connector list, and Muse builds it from the spec you point it at.

How often does the connector check for new mentions?

The connector checks whenever you ask Muse to, and RedReplier's own monitoring runs in real time. What the plan sets is the email alert floor: GET /alert-settings returns cadenceMinutes alongside minIntervalMinutes, which is 60 minutes on Starter and 15 on Business.

What happens when the RedReplier API changes after I build the connector?

Muse saves the connector as a reusable skill that holds whatever endpoints it learned the day it was built. No published source documents what Muse does when that API changes, so re-point it at the spec after any change.

Create a RedReplier API token and give Muse all four sources in one read-only call, or read the agent docs first.

Sources: Meta's Muse connectors help article, Meta's safety engineering post, Parallel's custom integrations write-up (14 September 2026), and the RedReplier OpenAPI spec. Checked 20 September 2026.