How Can an AI Agent Monitor Brand Mentions? MCP Tools, Sources and Real Costs

An AI agent monitors brand mentions with two abilities it does not have by default: a tool that searches public platforms for your name, and somewhere to remember what it has already shown you. Connect both through an MCP server and "who is talking about us?" becomes a question Claude or Cursor can answer with links, then answer again tomorrow with only what is new.

The agent's part is the reading. It can tell a customer complaint from an unrelated company with the same name, which is the step keyword alerts leave to you.

Which tools does an agent need to monitor mentions?

Two jobs, and they use different tools:

Job	Tool	What it does	Cost
"Who is mentioning us right now?"	find_mentions	Searches every source once and stores nothing	1 credit per source that answers
"Keep watching, and tell me what's new"	create_monitor	Saves the term and checks on a schedule, or when asked	Free to create; checks are charged
"What's new since I last looked?"	get_monitor_mentions	Reads stored results, only unread ones if you ask	Free
"Check again now"	check_monitor	Runs the sources immediately	Each source's normal price

Use find_mentions when you want an answer once. Use a monitor when the useful question is what changed.

What does a real mention search return?

Here is one call to the live server on 1 November 2026: find_mentions with the term Acme. It took about 10 seconds and charged 10 credits, about 1.25 cents at the $1.25 per 1,000 starter rate.

Source	Results read	Containing the term
YouTube	41	39
PeerTube	50	42
Bluesky	50	39
Dailymotion	50	35
TikTok, hashtag	29	29
Odysee	20	20
Hacker News	20	16
Instagram, hashtag	12	12
Threads	21	11
Reddit	7	1
Total	300	244

Three things in that table matter when you plan your own use.

The source report is the useful part. Each source returns its own outcome. "Reddit could not be reached" and "Reddit had nothing" are different answers, and an agent that gets them mixed up will tell you nobody is talking about you during an outage. All ten answered in this run.

A match is not yet a mention of you. Acme is a common word and a common company name, so expect many of those 244 to be about somebody else. We did not classify them, and neither does the tool: it checks that your term appears as a whole word, not which Acme is meant. That judgement is the agent's job, and it is why this suits an agent better than an inbox.

Hashtag sources work differently. Instagram and TikTok are searched by hashtag, so everything returned counts as a match, and you get the posts the platform shows for that tag rather than every post that uses it.

This is one sample of one term at one moment. It shows what a call returns and costs, not how much anyone is discussed on each platform.

How do you make the agent report only new mentions?

Create a monitor. The first check records what already exists and announces none of it, so you are not handed a backlog. After that, only items it has never seen count as new.

{
  "tool": "create_monitor",
  "arguments": {
    "term": "Acme",
    "schedule": "manual",
    "agent": "claude-code"
  }
}

We ran that call too. It set up all ten sources, returned creditsPerCheck: 10 and checksPerDay: 0, and its first check reported isBaseline: true. That check charged 0 credits, because the same searches had just run and a cached answer is free. A check with nothing recent to reuse charges up to the full 10.

With schedule: "manual" nothing runs until the agent calls check_monitor, so there is no recurring charge. Leave it out and the monitor checks every 6 hours by default, or as often as hourly. It is not a live feed, and a post can wait until the next check.

get_monitor_mentions with unread_only returns what this API key has not read yet, so the agent can begin any session with "what's new?" and pick up where the last one stopped. If you would rather be sent new mentions, a monitor can also take a webhook.

Two behaviours are worth knowing before you trust the first results. YouTube search and Instagram top posts carry no date, so results from those two can stay quiet for a day after the first check, and an older post can occasionally surface as new when a platform reorders. PeerTube and Odysee rank by match rather than date, so a new video on a busy term may not reach the results a check reads.

What does it cost to run every day?

Each check calls each source once, at 1 credit a source for these searches. A source that fails or returns nothing is free, and finding forty mentions costs the same as finding none.

Schedule	Most it can cost a day	About a month	Roughly, at $1.25 per 1,000
Only when the agent asks, once a day	10 credits	300 credits	$0.38
Every 6 hours, the default	40 credits	1,200 credits	$1.50
Hourly	240 credits	7,200 credits	$9.00

Those are ceilings for one term on all ten sources, and you can list fewer with platforms. Credits are sold in packs from $25 for 20,000, so these are shares of a pack rather than monthly bills. Your agent's model usage is separate, and reading a large result set takes tokens, so ask it to work through new items rather than everything each time.

Because a scheduled monitor outlives the session that made it, set an end with stop_at or max_credits. Monitors created with the agent field also have a 200-credit daily limit on projected spend by default, and that field is required over MCP.

Which sources are missing?

X, LinkedIn post search and Instagram captions. All three need a logged-in session or a paid search product, and these tools read only what a logged-out visitor can see, which is a line we do not cross. News sites, blogs and review sites are not covered either.

That gap decides this for many teams. If your customers talk about you on X or LinkedIn, use a tool that covers them.

How does this compare with a mention tool that has an MCP server?

The closest alternative is a monitoring product you connect an agent to, and Octolens is a good example. This comes from its own page, read on 1 November 2026, and pricing changes.

	Octolens	TrueScrape
Sources	X, Reddit, LinkedIn, YouTube, TikTok, Bluesky, Hacker News, GitHub, DEV, Stack Overflow, podcasts, newsletters, news and reviews	Reddit, Hacker News, Bluesky, YouTube, Threads, Dailymotion, PeerTube, Odysee, plus Instagram and TikTok hashtags
Pricing	Pro $159 a month, Scale $499 a month, billed annually; MCP on every plan, including the free trial	Per check from a credit balance; no monthly plan
What the agent reads	Mentions your plan has already collected	The result of a search it runs itself, or stored monitor results
Sign-in	OAuth, no API key	API key

Octolens is the better choice if you need X, LinkedIn or podcasts, or you want a dashboard your team lives in. TrueScrape suits an agent that checks occasionally, a small volume of terms, or a budget that should stop when the checks stop. Our comparison of brand monitoring tools covers the rest of the field, and tracking mentions without a subscription covers the same monitors without an agent.

What should you tell the agent to do with mentions?

Finding them is the cheap part. Most of the value is in the instructions. Something like this works:

Find mentions of Acme. Ignore anything about a different Acme. Group the rest by source, link every one, and flag anything that needs a reply. Say how many results you read and which sources failed.

Keep a person on the reply. A mention is text written by a stranger, and it can contain instructions aimed at your agent. Results from the server carry a notice saying they are data and not instructions, which lowers that risk without removing it. Drafting replies for review is a sensible use, and posting them unattended is not.

To try it, get a free key with 1,000 credits, run npx truescrape agent add claude-code, and ask who is talking about you. To check a name before connecting anything, the brand mentions checker runs a smaller version of the same search in a browser, with no account. Watching what a competitor posts is a different job, covered in how to give Claude access to competitors' social media.