Keep a Meta Muse Connector for Lead Finding on Digest Duty

Every other agent job fails quietly, but a Meta Muse connector for lead finding fails in public, because its output ends up in front of a stranger who never asked to hear from you. That one difference decides how you build it. The connector's job is to hand you a shortlist each morning. Choosing which thread to answer, and what to say, stays with a person.

The short version: let the connector read GET /mentions and call POST /mentions/{id}/explain, keep PATCH /mentions/{id}/status out of its reach, and set Connectors to "Always ask" under Settings then Permissions.

Why is lead finding the riskiest job to hand an agent?

Lead finding is the monitoring job whose output leaves your company: the reply lands in a stranger's thread under your account name. A brand monitoring agent that misreads a thread costs you thirty seconds of scrolling. A lead finding agent that misreads a thread gets a sales message into a subreddit where nobody invited one, and the account that sent it wears the result.

Meta's position on the mechanism raises the stakes rather than lowering them. The Meta Muse connector for Reddit monitoring guide covers how Muse builds one and where the credential goes. Meta's help centre says it "doesn't review custom connectors or how they use your information, so grant access with caution." Its guidance page adds that some agent actions cannot be undone and puts the work of fixing a mistake on you.

What do I paste into Muse to build the lead connector?

One message builds it, and the scope clause matters more than the spec URL. RedReplier publishes an OpenAPI document at https://redreplier.com/openapi.json that needs no auth to read, and the clause that makes it a lead finding connector rather than a general one is the list of calls you name.

Build a custom connector for RedReplier from the OpenAPI spec at
https://redreplier.com/openapi.json
 
Base URL: https://ai.redreplier.com/ai-app/api/v1
Auth: an Authorization: Bearer header. I will paste the token into the
secure credential prompt. Do not ask for it in the chat.
 
Allowed calls: GET /mentions, GET /mentions/count, and
POST /mentions/{id}/explain.
Never call PATCH /mentions/{id}/status. Never repeat the
aiReplySuggestion field back to me.
Your output is a list I read, nothing more.

Create the token at redreplier.com/api-tokens and paste it into the credential prompt, never the chat. The Reddit monitoring guide covers the token format, where Muse stores it and what a 401 means.

What does a daily lead digest call?

The digest is one GET /mentions call filtered by scoreBuckets and from, and the date filter is what makes it a digest rather than a re-read of your whole queue.

Every weekday at 08:30, list mentions added since yesterday's run,
HIGH and VERY_HIGH only, sorted by relevance score.
For each one: the source, the score, one line on what the person wants,
and the link. Explain the top three.
If there are none, say nothing.

A mention carries a relevanceScore from 0 to 100 or null, and the MentionSource enum in the spec covers REDDIT_POST, REDDIT_COMMENT, TWITTER, BLUESKY, HACKERNEWS, FACEBOOK and FACEBOOK_GROUP. So the same digest widens from Reddit to every platform RedReplier watches by changing the sources parameter, which is the cheapest way to test whether your buyers are somewhere you were not watching.

How do I check a relevance score before acting on it?

POST /mentions/{id}/explain returns the reasoning behind a mention's score, and it is the call that turns a number into something you can judge. The explanation tells you whether the person is asking for a recommendation, complaining about a competitor, or writing a thread you would be rude to walk into.

It is a POST because the spec describes it as re-running scoring and filling in relevanceReason, which is a write. It changes no mention status. Say that in your setup message, because an agent told "GET only" will skip the single most useful call available to it and hand you a digest of bare numbers.

Never act on a bucket alone. Two mentions sit in VERY_HIGH for the same keyword, and one is a buying question while the other is a two-year-old thread someone necroposted in. Our notes on lead scoring cover what a score with no reasoning attached is worth.

Which call has to stay behind a human?

PATCH /mentions/{id}/status approves or rejects a mention, and it is the call to keep away from the agent. A rejected mention leaves the queue, so a misjudged lead never appears again and you never learn that the agent made the call.

Should Muse ask before every action on this connector?

Set Connectors to "Always ask" for this one. As of 20 September 2026, Settings then Permissions offers two separate controls, Connectors and Web access defaults, and each takes either "Ask for some actions" or "Always ask". Engadget reported on 12 September 2026 that "ask for some actions" is Meta's default, which for Connectors means Muse asks before every write action and important read actions.

As of 20 September 2026 Meta publishes no per-endpoint permission surface for a connector it wrote for you, so the coarse control is the one you have. Run one digest and read it end to end before you trust it, checking that every link opens the thread the summary described.

What breaks on lead finding specifically?

The first failure is the digest boundary. "Since the last run" is the agent's own bookkeeping, not a field in the API, so a missed morning either double-reports yesterday's threads or skips them entirely. Ask for a fixed window, such as the last 24 hours, and accept a little overlap.

The second is the draft reply arriving unasked. Every mention from GET /mentions carries an aiReplySuggestion, and an agent summarising the response will hand it to you as suggested wording. Name the field in the scope, because the gap between reading a draft and pasting it is one tap.

The third is staleness in the saved connector. Parallel found on 14 September 2026 that Muse saves a custom connector as a reusable skill that persists into later conversations, and as of 20 September 2026 Meta's help pages cover disconnecting listed connectors without documenting how to delete one the agent built. Deleting the token at redreplier.com/api-tokens ends every call whatever state Muse is in, which makes it the kill switch to reach for first.

Frequently Asked Questions

Can Muse post a reply to a lead for me?

RedReplier's public API has no endpoint that publishes a reply, so the connector cannot post one however it is scoped. Driving a site in Muse's browser instead is a worse idea for lead finding, because unreviewed replies get accounts banned and the browser path breaks whenever the page changes.

What is the difference between this and the Grok Bot version?

The Grok bot template for lead finder runs the same calls behind the same redreplier_ token, and it is a shareable template rather than a connector Muse writes on demand. Its safety rail is billing, since keywords land as PENDING and cost nothing until activated. The Muse version's rail is scope, because Meta does not review what a custom connector does.

Does Meta Muse support MCP for this?

No MCP path is needed, because the connector reads the REST spec directly. The Reddit monitoring guide covers what Parallel found about MCP and the Muse consumer agent on 14 September 2026. For MCP-capable clients such as Claude, Cursor and Codex, RedReplier runs a hosted server at https://mcp.redreplier.com/mcp with six tools behind the same redreplier_ token.

Does the digest cost a RedReplier API call per mention?

One GET /mentions call returns the whole digest, and each POST /mentions/{id}/explain is one more. A digest of eight mentions with the top three explained is four calls. Every response carries RateLimit-Limit, RateLimit-Remaining and RateLimit-Reset, so the connector paces itself instead of finding the limit through a 429. The Reddit monitoring guide covers why a server on your own laptop is unreachable from Muse.

How many mentions should a daily digest contain?

Size the digest by what you will actually read rather than by a target count: filter scoreBuckets to HIGH and VERY_HIGH, and cap the call with limit, which defaults to 50 and tops out at 500. RedReplier's buckets are VERY_HIGH at 75 and above and HIGH at 50 to 74, so a digest that still runs long after that filter means the keyword set is matching too broadly rather than the market being hot. Tighten the keywords before you tighten the bucket. Buyer intent covers which signals justify a reply in the first place.

Can the connector cover Facebook and Hacker News, not just Reddit?

Yes. The MentionSource enum covers REDDIT_POST, REDDIT_COMMENT, TWITTER, BLUESKY, HACKERNEWS, FACEBOOK and FACEBOOK_GROUP. Change the sources parameter and the same digest widens to every platform RedReplier watches.

How do I shut the connector down quickly?

Delete the token at redreplier.com/api-tokens. Every call ends whatever state Muse is in. Meta's help pages cover disconnecting listed connectors but do not document deleting one the agent built, so the token is the route you can rely on.

What do the HIGH and VERY_HIGH buckets mean?

VERY_HIGH is a relevance score of 75 and above, and HIGH is 50 to 74. The score runs from 0 to 100, or is null. A bucket alone does not tell you whether the thread is worth entering, so run POST /mentions/{id}/explain on the ones you care about.

Should I paste the aiReplySuggestion text Muse shows me?

Treat it as a draft you read, not wording you send. The reply lands under your account name in a stranger's thread, and the gap between reading a draft and pasting it is one tap. Name the field in the connector's scope so the agent does not repeat it, and write the words yourself.

Does Meta check what a custom Muse connector does?

Meta's help centre says it doesn't review custom connectors or how they use your information. That leaves two rails you control. One is the list of calls in your setup message, and the other is Connectors set to "Always ask" under Settings then Permissions.

Create a RedReplier API token and give Muse a read-only lead digest tomorrow morning, or read the agent docs first.

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