Claude (MCP) vs. KOLens Web UI — When to Use Which (and What Each Costs)

Where the confusion comes from

A recurring piece of user feedback after we shipped the MCP plugin: "I tried Claude with KOLens MCP but Claude itself charges for AI tokens on top of KOLens credits. When does that pay off versus just using the UI? I'm confused about the cost model."

Fair question, and the honest answer has two halves. The KOLens bill does not change with the interface — a search is a search. The Claude bill is separate and depends on which Claude tier you are on. We want to stop being cute about this in our marketing and write down the actual decision matrix.

The two interfaces, in plain terms

KOLens exposes the same core data and actions through two surfaces. They share one underlying account, one credit balance, one workspace, and the same authentication.

1. The KOLens Web UI

Open kolens.ai, log in, and you get the full product: keyword search, Discovery Plans (recurring auto-scrape), Watchlist (monitoring with growth and dormancy signals), KOL Lists (outreach pipeline), the find_similar_creators lateral-expansion tool, Excel and CSV export, and AI-drafted outreach. Everything is point-and-click with structured forms, sortable tables, and visible credit balances.

2. The KOLens MCP plugin in Claude

Install once via /plugin (or connect at /connect for the hosted MCP). Claude gains two skills:

• kol-search — quick chat-style lookups against the accumulated KOLens database. No credit cost beyond what a normal search costs.
• kol-research — a deeper run that scrapes up to 200 creators for a query. Costs 1 KOLens credit per run.

Both skills run inside a conversation, so you can chain them with whatever else Claude is doing — reading a brand voice file, querying Notion, posting to Slack, drafting an email.

The real cost breakdown

Two bills, never combined. Here is what each side actually charges.

KOLens side (same for both interfaces)

• Starter $29 / 50 searches — works out to $0.58 per search. Best for light users figuring out the workflow.
• Growth $79 / 200 searches — $0.40 per search. Best for steady weekly use.
• Pro $149 / 500 searches — $0.30 per search. Best for agencies and ops teams running multiple brand accounts.

Credits never expire. A search is a search whether it came from a chat prompt or a button in the UI — the line item on your KOLens invoice is identical. See the full table at /pricing.

Claude side (only when you use the MCP plugin)

• Claude Pro at $20/month — a flat subscription with generous fair-use limits. Once you are paying this, the marginal cost of each KOLens conversation is effectively zero on Claude's side. Almost always the right tier if you talk to Claude daily.
• Claude API, per-token — roughly $3-15 per million input tokens and $15-75 per million output tokens depending on the model (the exact rates change; check Anthropic's pricing page for current numbers). A six-minute multi-tool run typically lands between $0.20 and $0.60. Better for engineering teams building automations than for daily interactive use.

What a single search actually costs

When Claude (MCP) wins

The interface buys you three things the UI cannot: a chat as the input surface, an LLM doing the filtering and ranking, and free composition with every other MCP server you have connected.

Ad-hoc lookup mid-conversation

You are on a call with a brand partner who name-drops @somehandle. You want to know follower count, recent posting cadence, and whether they have a contact email — without switching tabs. Type the question into Claude. kol-search fires, you get the answer in the same window, and the conversation continues. The UI would require five clicks; Claude requires one sentence.

AI-driven filtering and ranking

"Find me 5 phonecase TikTok creators with email plus at least 50K followers, audience skewing US or UK, then pick the top three by posting cadence and ICP fit." That last clause is the Claude advantage — the UI returns the raw filtered list, but Claude can read each candidate and apply qualitative judgment before handing you a shortlist. For nuanced briefs where the right creator depends on more than numeric filters, this is the difference between getting 30 names and getting 3 good ones.

Chaining KOLens with other MCP servers

The cross-platform composability is where MCP earns its keep. Connect KOLens, Notion, Slack, and your filesystem in one Claude session. A single prompt can: discover phonecase creators through kol-search, dedup them against your Notion partnerships database, draft personalized outreach using a brand voice file from the repo, write the drafts to local markdown files, and post a summary in Slack. No KOLens UI can do that — we do not own Notion, Slack, or your filesystem. See the case story at find-kols-with-claude-code-mcp.

Async-friendly workflows for cross-border sellers

Many of our 出海 (cross-border) sellers work async with their TikTok managers across time zones. A quick "is @somehandle worth looking at?" question dropped into Claude returns a useful answer in seconds without booking a synchronous review. The chat log itself becomes the handoff document — the manager picks it up the next morning, sees the same answer Claude gave you, and keeps moving.

When the Web UI wins

The UI is built for the workflows that have state — pipelines, schedules, lists, monitoring. Conversational interfaces handle those poorly because the UI of a chat is just text, and structured data wants structured controls.

Bulk operations on lots of rows

Sorting a 200-row search result by engagement rate, multi- selecting 40 of them, adding all 40 to a KOL List, then bulk- tagging them with a niche label. This is three minutes in the UI and a frustrating back-and-forth in chat. When the operation is "do the same thing to many rows", point-and-click wins decisively.

Discovery Plans setup

Discovery Plans (recurring auto-scrape jobs that grow your creator library while you sleep) involve a structured form: keyword, schedule cadence, filter set, region targeting, role classification. The UI has dedicated forms with sensible defaults. Setting one up via chat means dictating each field — slower and more error-prone. Read more at automated-tiktok-kol-discovery-plans.

Watchlist tracking with signal toggles

Watchlist gives you per-row badges across six signal kinds (rising_kol, trending_video, dormant, new_sponsored, new_contact, no_outreach) and an at-a-glance digest of windows.{7d,30d}deltas. That is a visual surface — the chat version would have to read out tables row by row. Open the Watchlist page; do not try to navigate it from Claude.

KOL Lists pipeline (the outreach funnel)

KOL Lists are the decision surface for who you actually contact: fit-ranked rollup, triage state per row, CRM-derived outreach history. Moving a creator from "candidate" to "contacted" to "negotiating" to "live" is a status update that the UI handles with one click and a dropdown. The chat equivalent is a round-trip per row.

Excel and CSV export for handing off to a team

Sometimes the next step is "send this spreadsheet to the partnerships lead in Shenzhen". The UI has a one-click export to .xlsx; Claude can dump JSON but not produce an actual formatted spreadsheet with sortable columns. When the deliverable is a file, the UI is the right tool.

Predictable per-search cost

Finance teams ask us this one a lot: "I want to budget $300 a month for KOLens." From the UI, the math is trivial — 750 searches on Pro, hard cap. From Claude, you also have to track the Claude bill on top, and a verbose conversation can rack up Claude tokens faster than you expect. If predictable per-line- item cost matters, the UI keeps the math clean.

Decision matrix — pick by use case

Concrete scenarios. Pick the row that matches what you are actually trying to do.

The hybrid pattern most teams settle into

After watching how customers actually use both surfaces, a clear pattern shows up. The UI handles the structured, stateful, long- running pieces. Claude handles the conversational, ad-hoc, cross- tool pieces. They are not competing — they cover different parts of the same workflow.

1. 1

   Set up the durable infrastructure in the UI.

   Configure two or three Discovery Plans for your core niches. Add your most-watched 30-50 creators to the Watchlist. Create a KOL List per active brief. This takes an hour, lives forever, and runs in the background without further chat.
2. 2

   Use Claude for the daily ad-hoc layer.

   When a creator gets mentioned, when a brief evolves, when you want a qualitative shortlist over a numeric filter — drop the question into Claude with KOLens MCP connected. You are answering a question, not editing a database.
3. 3

   Compose Claude across your stack.

   Connect KOLens alongside Notion, Slack, Gmail, your filesystem, and the Cowork plugin. One prompt now spans discovery, CRM updates, and team notifications. This is the workflow the UI structurally cannot do.
4. 4

   Come back to the UI for the deliverable.

   When the chat run is done, the rows live in Notion or KOLens Lists. Open the UI to export to Excel, refine the triage states, send a polished outreach. The chat got you to a decision; the UI ships it.

Roughly speaking: 70% of our active users have both surfaces connected. The MCP plugin install does not commit you to one side or the other. It just gives you the option.

Getting started with either

Both surfaces start from the same account. Sign up once, get 50 free searches on the Starter feel-it-out tier, then pick the interface that matches the workflow you actually need.

• Web UI: /register creates the account and drops you on the search page. From there, Discovery Plans and Watchlist are one nav click away.
• MCP plugin: /install walks you through connecting KOLens to Claude (or any MCP client). The hosted setup is at /connect; see the OAuth walkthrough at connect-kolens-to-claude-mcp-oauth.
• Direct discovery in the UI: /find-tiktok-influencers runs a no-signup search if you want to see the product before paying anything.

Building the creator library either way

The compounding workflow — Discovery Plans accumulating a library of creators week after week — is UI-native and we wrote about it in detail at build-tiktok-creator-library-discovery-plans. The Claude layer sits on top of that library, querying it with natural language. If you have not built the library yet, start in the UI; if you have, layer Claude in for the conversational questions.

Frequently asked

Does using Claude make each KOLens search more expensive?

Not on the KOLens side — a search costs the same credit whether you trigger it from Claude or from the web UI. The extra cost is on Claude's side: every message in the conversation consumes Claude tokens, either covered by Claude Pro at $20/month flat or billed per-token through the Anthropic API.

Which interface is cheaper for light usage?

If you already have Claude Pro for other reasons, adding KOLens via MCP is essentially free on the chat side. If you do not have Claude Pro and would be paying for chat tokens just to access KOLens, the web UI is strictly cheaper — your only line item is the credit itself.

Can I run Discovery Plans and Watchlist through Claude?

You can list and inspect them via MCP tools, but creating and editing them is awkward in chat — those are UI-native workflows with structured forms and visual digests. Most teams set them up once in the UI and use Claude for the ad-hoc lookups in between.

What about Cowork compatibility?

The KOLens MCP plugin is Cowork-compatible, so the same install works in Cowork's environment alongside Claude. The pricing model is identical — KOLens credits per search, plus whatever Cowork or Claude charges for the chat side. The decision matrix in this post applies equally to both clients.

Where do I see exactly what each tool call costs?

The KOLens dashboard at /account/usage shows every search, every credit charge, and the source (UI versus MCP). On the Claude side, the usage console shows per-conversation token spend if you are on the API; Claude Pro is flat-fee so there is no per-conversation line item. Cross-reference the two if you want to attribute spend by interface.