2,400 tools through one MCP connection. Agentik.

10/12/26 on mcp, ai, tooling

MCP has quietly become a package-management problem. Every new capability for your agent means a new MCP server: find the repo, check whether anyone still maintains it, install it, get an API key from the provider, paste a config block, restart the client, hope. Do that five times and your MCP config reads like the dependency file of an abandoned side project — ten entries, half of them stale, and a key-rotation ritual that touches five places.

There is a deeper issue underneath the sprawl. The coding loop is already covered: shell, git, a filesystem, code execution. What an agent genuinely cannot do lives behind fifty other APIs — post to social, pull ad spend, check search rankings, scrape a page, generate a video. And nobody wants to babysit fifty MCP servers to get there.

What it is

Agentik is a tool gateway for agents: one MCP connection in front of roughly 2,400 tools from about fifty providers. You connect your client once — Claude Code, ChatGPT, Codex, Cursor, anything that speaks MCP — and the whole catalogue appears as a single server. Agentik routes each call to the provider API behind it and bills per call, at the provider’s own price, from one prepaid balance. No markup, no subscription per tool.

The connection

The onboarding is the part I enjoyed most, because it is written for the agent rather than for you. You tell the agent to read https://www.agentik.cc/connect.md and connect itself:

$ claude
> Read https://www.agentik.cc/connect.md and connect yourself to Agentik.
  ⏺ Reading the page: server URL, auth steps, what to verify…
  ⏺ Added the MCP server. Listing tools…

One entry in the MCP config instead of a paragraph per tool. Whoever decided that the setup document should be machine-readable Markdown that the agent consumes directly understood the audience: if the pitch is “your agent does the work”, the setup should be something your agent can do.

The shelves

It is a marketing-and-growth catalogue, and it does not pretend otherwise:

Social. TikTok, Instagram, YouTube, X, Reddit — publish, schedule, respond to comments, watch mentions.
Ads and analytics. Meta Ads, Google Ads, TikTok Ads, LinkedIn Ads, plus GA4 and Search Console — spend, creative performance, search queries, on demand from a prompt.
Search and web. Web search, SERP and SEO data, AI-visibility tracking (where a brand turns up in AI answers), scraping, Maps data, browser automation.
Business data. B2B lead databases, Amazon and e-commerce listings, app stores, reviews, finance data.
AI media. Video, image and audio generation, with fal.ai underneath.

The point of that list is not its length. It is that every line on it used to be its own MCP server with its own install, key and config block. Collapsing the whole shelf into one connection is the entire product.

Playbooks

There is a library of ready-made recipes — a bit over a hundred of them. An AI-visibility audit (which brands an AI assistant names for a query, and whether yours is among them). A weekly marketing report assembled from GA4 and the ad platforms. A competitor ad teardown. A B2B lead list from a plain-language description of your customer. They are prompts, not code: paste one, the agent picks the tools and assembles the result. As a way to learn what the catalogue is actually good for, this beats reading a feature grid.

Pricing

Per call, at the provider’s own price, out of one prepaid balance. No subscription per tool, no seat maths, and nothing added between you and the provider’s rate. The comparison to keep in mind: a subscription caps your cost, metering does not. More on that below.

The honest trade-offs

It is a middleman. One more service between your agent and the provider APIs — a dependency and a trust decision, and worth saying out loud. If the gateway is down, the agent still runs; it just cannot post the tweet.
Metered billing has no ceiling. The subscriptions you skipped would have capped spend; per-call does not, and the agent decides how many calls “check our rankings” means. Watch the balance the first week and set your own limit before the agent develops opinions about thoroughness.
The provider accounts are still yours. Write operations — publishing, scheduling, responding — go through accounts you connect and permissions you grant. Agentik is the plumbing, not a social media manager; approvals and revocations remain your problem.
The catalogue is growth-shaped. If your agent’s weekend involves kubectl or tofu plan, this is not that shelf. Ops tooling still belongs to the usual MCP servers.
Aggregation is not improvement. A gateway can route to a provider’s API; it cannot make a flaky one less flaky. Provider quirks survive the trip.
It is young and moving. A tool count that climbs weekly is healthy, and also a reminder that the catalogue page, not this post, is the source of truth.

Where I landed

Two weeks ago the post here was about running one agent per branch, because the working directory had become the bottleneck. The other half of that loop — what the agent can reach beyond the filesystem — has been the messier half for a while. Agentik is the tidiest answer to it I have seen: one connection, one balance, and a catalogue covering the boring external world of social, ads, search and leads that used to sit fifty config blocks away. My MCP config went from a graveyard to one line. That is the whole recommendation.