My mention monitor charges per search, not per month

Every small product eventually wants to know who is talking about it. Someone asks on Reddit whether your tool is worth it. A thread on X gets your pricing wrong. A YouTube review goes up and you hear about it a month later from a customer.

The established tools solve this for marketing departments. Octolens starts at $199 a month ($159 billed yearly) and Brand24 at $249, with a dashboard built for someone whose job is the dashboard. For a product with a few hundred users that is five to twenty mentions a week, read once a day by the person who also writes the code. The price is set for the wrong buyer.

I built Overhear for the other buyer. It is mine, so calibrate for bias.

One credit is one search#

The cost of mention monitoring is not the user. It is the search: every check of every source is a vendor call, and every hit is a model call. So that is what Overhear charges for.

A credit is one search of one keyword on one source. Reddit, the web and Hacker News are one credit, X and YouTube two, LinkedIn three, roughly in proportion to what each costs me. Credits come in packs ($10 for 6,000, $25 for 16,000, $50 for 35,000), they never expire, and when they run out scanning pauses until you top up. No subscription, no seats, no keyword tiers. The first 1,000 are free, so the first scan runs before you pay anything.

Every keyword has one setting per source: hourly, every 6 hours, daily, or off. A brand name at the default pace (X and Reddit hourly, LinkedIn every 6 hours, YouTube, the web and Hacker News daily) costs about 2,640 credits a month, which is $4.40 on the small pack. Everything once a day is 300 credits, 50 cents. The Usage tab turns the schedule into a burn rate and says how many days the balance lasts. You get one email at seven days and one at zero.

What a scan does#

A scan runs the keyword through the source’s search, keeps only posts that actually contain it (vendor search is loose, and Reddit happily returns posts that never say the word), drops your own accounts, and stores each post once per keyword. New posts go to Claude Haiku with the project description and come back with one type: buying intent, complaint, bug report, question, praise, misinformation, competitor, or other, plus a sentiment. You can add up to ten categories of your own with a one-line description. “Pricing: talks about what it costs” beats the built-in types when it fits.

Sorting learns from the team. “Not related” and “Related” clicks become examples in the next classification of that keyword, so the verb “overhear” stops showing up after you dismiss it a few times.

Misinformation is the type I wanted most. You confirm a short list of facts about the product (pricing, plans, platforms, what it does not do) and any post that contradicts them is flagged with the claim and the fact side by side. Once the facts are confirmed, a weekly sweep searches the web for “yourproduct vs”, “yourproduct alternatives” and “yourproduct pricing”, reads up to ten pages and checks those too. The listing site that still shows last year’s price becomes a mention with a drafted correction request to its owner.

Replying#

A mention you can answer has a Draft reply button. It is 15 credits a click and runs on Sonnet: under 70 words, one concrete tip first, a plain “I work on Overhear”, and on Reddit no link, because automod removes links from low-karma accounts and never tells you. The first version of that prompt asked for two to four sentences and got four 40-word sentences, a tutorial with the product at the end. A word budget and two example replies fixed it. A longer rule would not have.

Nothing is posted for you. The draft is text you paste, and Mark replied records that you did.

Ask it instead of reading it#

A feed is still a thing you have to open. The part I use most is the connector.

Overhear runs an MCP server at /mcp with OAuth. In Claude you add it as a custom connector, sign in once, and ask:

Which mentions need a reply today?

Save this week’s praise as testimonials.

Which posts got our pricing wrong this month?

The tools cover the feed, the reviews (related, unrelated, testimonial, replied), keywords and their pace, and drafting a reply. A skill, overhear-mentions, tells the agent what to check first, what to leave to you, and which tools cost credits.

I wrote earlier this year about why SeggWat got an MCP server. The argument is stronger here. A mention feed is read by one person, once a day, with a question in mind. That is a chat, not a dashboard.

The Rust side#

After a year of Dioxus fullstack I wanted the opposite: server-rendered HTML, no WASM, no 90-minute builds. Overhear is axum with Topcoat for the views, Postgres through sqlx, and Tailwind compiled by a build script from the oxidt-ui component classes. The whole app ships about 16 kB of JavaScript: passkeys, the theme toggle, PWA registration. Charts are inline SVG built in Rust. Settings are plain form POSTs with a redirect and a flash message, so the dashboard works with JavaScript off.

The domain lives in its own crate, radar-core: which vendor endpoint answers each source, how a response becomes mentions, the prompts, the credit weights. No sessions, no OAuth, no rendering. The app crate is a monolith around it. I skipped that split with SeggWat, and its dashboard crate is now 99k lines with a 22-minute WASM compile per commit, so this time it happened on day one.

Searches go through treg, a tool gateway that routes each source to the vendor with the best measured hit rate and falls back when one fails. Self-hosting is a Docker image and a Postgres. A self-hosted instance has no credits; its Usage tab shows what the vendors and the models actually cost.

What it is not#

• It polls. There is no firehose. Hourly is the fastest pace, so a post shows up within the hour, not the minute.
• It finds what search finds. Reddit’s own search is loose, and some posts never come back for a word they contain.
• The UI is English only. A German catalog existed and translated the shell but not the content, so it went.
• Nothing is posted for you. Replies are drafts you paste. That is on purpose.

Try it#

Sign up at overhear.dev, type a domain or a keyword, and the first scan runs on the free credits. The connector setup is in the MCP guide. If a mention gets the wrong type or a draft reads like an ad, tell me. The reply prompt changed this week because of exactly that.