Lifecycle Part 5: AI + MCP for KOL Research at Scale

The first four parts of this series walked the entire manual lifecycle: discovery, vetting, library, monitoring. Done well, the four stages compress a 40-hour-per-launch workflow into roughly two hours. This final chapter is about giving even those two hours back.

Stage 5 is the leverage chapter. Once the workflow is in shape, you can hand most of it to Claude — Claude Desktop or Claude Code, your pick — using the KOLens MCP server. The prompts that used to be five clicks across three pages become a sentence. The Monday digest that used to need a human reader becomes a Slack message Claude summarised. The 32-creator outreach pass that used to take a marketer half a day becomes a 6-minute background job. None of this is theoretical; we have a written case story from a 4-person phonecase brand doing exactly this in production.

Why MCP changes the shape of the work

The Model Context Protocol — MCP — is Anthropic's open spec for connecting LLMs to external tools and data. An MCP server exposes typed tools; an MCP client (Claude Desktop, Claude Code, Continue, Cursor, OpenAI's agents, anything that speaks the protocol) discovers and calls them. The protocol is simple. The implication for influencer marketing is large.

Before MCP, an LLM could read your dashboard's HTML if you pasted it in. It could write outreach copy if you described the creator. What it could not do was call the dashboard's underlying actions — search for creators, score authenticity, save a watchlist, draft outreach personalised to a real audience snapshot. The integration story was a long tail of zapier-ish webhook hacks and copy-paste loops.

With MCP, the integration story collapses. KOLens exposes about 25 tools at https://kolens.ai/mcp. Claude (or any compliant client) calls them like any other tool — typed inputs, typed outputs, OAuth-scoped permissions, no glue code. The workflow you used to click through, Claude can drive.

Setup — one URL, one click

The setup story has deliberately one moving part. From Claude Desktop:

1. 1

   Open Claude → Settings → Connectors → Add custom connector.

   The Connectors panel ships in current Claude Desktop and Claude.ai versions. If you do not see it, update the client.
2. 2

   Paste https://kolens.ai/mcp and click Connect.

   KOLens speaks the MCP streamable-HTTP transport with OAuth 2.1 + PKCE and dynamic client registration. Claude handles the protocol details.
3. 3

   Log in to KOLens and click Approve.

   Default scopes are read:kols + write:lists — enough for discovery, vetting, watchlist mutations, and outreach drafts. The third scope, run:scrape, is not requested by default so a careless prompt cannot spend credits without your explicit consent.
4. 4

   Start prompting.

   Claude now has the KOLens toolset. Test with "What can you do with KOLens?" — Claude will list the tools it discovered from the server.

Step-by-step screenshots and the OAuth flow detail live in the connector OAuth setup post. For the desktop-vs-Claude-Code decision see MCP vs web UI.

Real prompts that actually work

The point of MCP is not "Claude can call list_kols". The point is that Claude can compose tools across the entire lifecycle in one turn. Below are the prompts that actually produce useful work, in increasing order of leverage.

Prompt 1 — vetting a single handle (the simplest case)

Look up @somehandle on KOLens. Are they dormant? What is their authenticity score? Should I reach out?

Claude calls get_kol_profile, reads the cadence + authenticity + audience fields, and answers in prose. This is the entry-level use — fast, free, replaces a click-through to the dossier when you just need a verdict.

Prompt 2 — pulling a vetted cohort

Find 20 mid-tier US fitness creators on TikTok with engagement rate above 5% and a US audience over 60%. Give me a ranked list with emails.

Claude calls list_kols (or run_scrape if explicitly scoped), filters the results inline against your criteria, formats a table, and returns it. The same workflow takes 20 minutes of clicking in the web UI; Claude does it in 90 seconds and produces a copy-pasteable table.

Prompt 3 — end-to-end with outreach drafting

Find 20 mid-tier US fitness creators. Vet them for authenticity and audience country. Add the survivors to my Q3 watchlist. Draft personalised outreach for the top 10 using my brand voice from the doc I shared.

Claude calls list_kols + get_audience_snapshot per candidate, filters, calls add_to_watchlist for the survivors, then drafts personalised outreach using the brand voice document already in the conversation. The full pass — discovery, vetting, library write, outreach drafts — completes in roughly six minutes. The phonecase brand we have written about elsewhere reports this prompt replaces about a full afternoon of manual work per launch.

Prompt 4 — monitoring summary

Read my watchlist. Tell me which signed creators are showing silent-quit signals, which watchlist candidates are trending this week, and which competitors have signed creators we were tracking. Format as a Monday digest.

Claude calls list_alerts + list_watchlist, groups the signals into the three buckets you asked for, and returns a formatted Markdown digest. Paste into Slack or pipe to the Slack-MCP server and Claude will post it directly.

A worked example — six hours becomes six minutes

The phonecase brand from earlier chapters wired Claude Code to KOLens, Notion (their existing CRM), and Slack (their team channel). One prompt now runs the whole loop:

Find 5 US phonecase TikTok creators with engagement rate above 6%, audience US over 70%, no dormant signal. Add them to the Q2 watchlist. Log them in the Notion partnerships database with brand-fit notes. Draft outreach in our brand voice and post a summary in #partnerships on Slack.

Claude calls the KOLens tools to find and vet the candidates, the Notion MCP server to write the rows into the brand's existing partnerships database, and the Slack MCP server to ping the team. End-to-end time: roughly six minutes. End-to-end cost: about $0.40 in Claude usage plus 1 KOLens credit. The previous workflow — half a day every Friday — is gone.

The full case story (with prompts, screenshots, and the OAuth flow) is in Find KOLs with Claude Code + KOLens MCP.

When MCP wins and when the web UI wins

MCP is not the answer to every task. The honest split:

• Use MCP when the task is repeatable, natural language describes it faster than clicking, the output is structured (a list, a table, a draft), or you want to compose KOLens with another tool (Notion, Slack, your CRM). The end-to-end launch prompt above is the canonical case.
• Use the web UI when you want to see visualisations (sparklines, dossiers, the watchlist layout itself), browse exploratively without a clear query, or work with a teammate looking at the same screen. Discovery is often UI-first; outreach is often MCP-first.

The deeper decision-guide is in MCP vs Web UI — when to use which.

Where KOLens enters the workflow

• /connect — the OAuth setup surface for adding KOLens as a Custom Connector to Claude. One-time login per client.
• /plugin — the bundled Cowork plugin (stdio MCP) for environments where the HTTP connector is not available. Subset of tools but same auth model.
• /openclaw — the integration surface for the open agent ecosystem; for teams wiring KOLens into agents beyond Claude.

Why this is the future of influencer-marketing operations

The pattern across the case stories is consistent: once a team has the manual lifecycle running well, the MCP layer compresses the human-time portion of every launch by 80-90% without losing any control. The marketer is still the decision-maker. The marketer just stops being the click-through layer.

Two structural reasons this trend is durable.

First, the work in influencer marketing is repetitive in shape but bespoke in content. Every launch has the same five stages; every launch has a different niche, target country, brand voice, and constraint set. That mismatch broke traditional automation (too rigid) and burned out humans (too repetitive). LLMs handle exactly this shape — the structure is in the tools (MCP), the content is in the prompt.

Second, the cost curve favors the model layer over the human-hour layer indefinitely. A 6-minute Claude run cost the phonecase brand $0.40 in Q1; the equivalent run today on a newer model costs less. The same marketer hour has not gotten cheaper. The economics of the workflow tilt further toward "Claude drives, the human reviews" with every model release.

Teams that adopt this pattern now are not adopting a productivity hack. They are restructuring the unit of labour for influencer-marketing operations. The teams that wait will be competing against teams running ten launches for the cost of their one.

vs the alternatives

Closing the loop — the lifecycle as one prompt

Looking back across the five parts: discovery picks the funnel inputs, vetting cleans the funnel, the library compounds the asset, monitoring catches what changes, and AI / MCP collapses the human-time cost of running the whole thing. Done together, the lifecycle is the difference between treating influencer marketing as a series of one-off projects and treating it as a system that accumulates value.

If you are new to KOLens and reading this end to end, the recommended starting move is to run one search on /search, save the result to a list, and then come back to this chapter and wire Claude. The whole lifecycle becomes legible in about an hour of hands-on work. The compounding starts the same day.

The full lifecycle series

1. Part 1 — Discovery: Finding TikTok candidates at scale
2. Part 2 — Vetting: Separating signal from noise inside the funnel
3. Part 3 — Library: Turning finds into a durable asset
4. Part 4 — Monitoring: Tracking the picks you made
5. Part 5 — AI / MCP (you are here): Claude integration for natural-language research