Agent skills are live — 8,089 listed, imported from public repositories with each SKILL.md on its listing page. See what a skill is
Publish Sign in

Tools for your agent.
Revenue for whoever built them.

A marketplace for MCP servers and agent skills: publishers list them and get paid, and you subscribe to what you need — free, or by the month. Install any of it in one command into whatever AI tools and runtimes you already use; we handle OAuth, metering, one invoice and the payouts to publishers.

Try
Scanned on publish, and the grade is public Pin & roll back any version A 5xx or an unreachable server is never billed
One command, any client 19,800 servers · 8,089 skills
npx mcprush@latest add <server>
Works with6 of 15 clients by command, the rest by key
Billingone key, one invoice, a cap you set
Not billedrefused here, unreachable, or a 5xx
Free to start19,749 of 19,800 servers
Verified publishers1 of 16,848
19.8k
MCP servers
ready to install in one command
8,089
Agent skills
instructions an agent loads on demand
20
Stacks
sets that already work together
16.8k
Publishers
3 of them are earning
Catalogue

What is being pulled from the MCP marketplace right now

The first tab is the last twenty-four hours of package downloads, counted by npm and PyPI rather than by us. Best free is the same figure over the whole life of the package. Highest trust is our scanner's letter first, and among servers holding the same letter the one people pull most — and only servers that actually publish tools, because a server with none is a grade about nothing.

Browse all 19,800
AR
Arxiv
blazickjp·Research

A flexible arXiv search and analysis service with MCP protocol support

no installs yet 19 tools
Free
TI
Timeline Pulse
davidmosiah·Memory & Knowledge

MCP server + public atlas of timeline-shift folklore, Mandela Effect reports, and collider-era myth — every claim labeled, every source

no installs yet 10 tools
Free
CL
Claude Video Vision
jordanrendric·Design & Media

MCP server that gives Claude Code the ability to watch and understand videos — extracts frames via ffmpeg and processes audio via multiple

no installs yet 6 tools
Free
Browse by capability

Twenty-eight categories, one permission model

Agent skills · new

A server gives your agent hands.
A skill tells it what to do with them.

An agent skill is a folder of instructions — a SKILL.md and whatever reference files it needs — that loads into the context window only when the work matches. It changes how the model decides. It does not change what the model can reach, which is still the server's job.

Browse all 8,089 skills
Loads on a trigger

You type “review this PR before I push” and the instructions arrive. On every other turn they are not in the context window at all, so an installed skill you never invoke costs you nothing.

Priced in context first

The listed skills weigh between 25 and 1,067,300 tokens. That is the number the catalogue filters on, and the budget tray totals a selection before you install any of it. 6,128 of the 8,089 are open source; seventy-six of them are paid once — there is nothing on a skill to meter, and one price buys the folder and every update its author ships inside it. The files stay on your disk whatever happens next.

Composes with servers

461 of the 8,089 call an MCP server to do their work, and the card names how many. The other 7,628 are instructions only — no install and no scopes, though a skill can still carry a price.

Four free ones from four publishers you already know, picked as examples of a skill written out in full — the folder, the trigger, the licence and what it weighs. Every other skill in the catalogue is filed the same way.

CA
Canvas Design
anthropics · Output format

Create beautiful visual art in .png and .pdf documents using design philosophy.

No reviews yet no installs yet 35.5k ctx standalone
Free Details
HU
Hugging Face LLM Trainer
huggingface · Workflow

Train or fine-tune language and vision models using TRL or Unsloth with Hugging Face Jobs infrastructure.

No reviews yet no installs yet 46.6k ctx standalone
Free Details
CO
Connect Recommend
stripe · Output format

Use this skill when the user asks about Stripe Connect configuration, charge patterns, Dashboard access, or how to get started with Connect, is building a marketplace, platform, multi-vendor store, gig platform, or subscription platform, needs to pay out sellers, vendors, or providers, mentions spli

No reviews yet no installs yet 35.9k ctx standalone
Free Details
Works everywhere

Install once. Your whole toolchain sees it.

mcprush speaks Streamable HTTP and stdio, so the same server drops into every MCP client you use. Team installs roll out to everyone's config automatically — and revoke just as fast.

CCClaude CodeCDClaude DesktopGPChatGPTCUCursorVSVS CodeCXCodex CLIGMGemini CLIGKGrokMSMicrosoft CopilotPXPerplexityDSDeepSeekZEZedWSWindsurfSDAgent SDKAPHTTP / API
Team rolloutone file
# one file, committed to your repo servers: - id: <server>@<version> budget: $40/mo approve: [<tool>] - id: <server>@<version> scopes: [read] - id: <server>@<version>
Versions are pinned by default. A publisher cannot change a tool description under you — pinned installs keep the manifest you approved.
Trust layer

The risk of a tool is not its code. It's its reach.

A listing published here is scanned on publish and re-scanned on every release, and its report is on its own page. A listing imported from a public index is scanned from that public source instead, and its grade says which release it covers; where there is nothing to read, the page says so rather than inventing a letter. What you approve is a capability budget — not a promise.

Capability scopes, in plain language

"Reads your repo", "sends email on your behalf", "can reach any host". Declared in the manifest, enforced at the gateway, shown before install.

Network reach, named before you install

A listing's tools are labelled read, write or network from its manifest, and the network ones are named on its page. We do not yet hold a per-host allowlist, and the gateway does not refuse a call by its destination.

A scan report for every release

Tool descriptions are prompts. Every release gets its own scan report, with a digest of that release's tool surface and the tools it declares — so a surface that moved leaves a different digest. Warning you when it does is not built yet.

Pin, diff, roll back

Installs resolve to an exact release, fingerprinted by the scan of that build. Pin or move an install to any published version of a listing from its own page.

Install review · google-sheet-mcp · 2.1.2Interactive example verified: repo
Read the pages you ask for
sheets_list_tabs, sheets_read_range, sheets_get_sheet, sheets_write_range, sheets_create_tab, sheets_append_row, docs_create, docs_read, docs_get, docs_write, docs_replace, docs_list
read
Static analysisA0 findings
Verificationrepo
Prompt-injection surfaceno tool returns untrusted text
Open the listing Nothing is installed from this page
For publishers

Charge for your server without building a billing system.

Ship the tools. We do identity, rate limits, plan allowances, invoices, dunning and tax; Stripe Connect holds your bank details and pays you. A paid listing runs on your own infrastructure and reaches buyers through our gateway, so an expired subscription stops working without you writing a line of billing code — or list it free and keep the distribution.

FreeList it openly, source stays yoursSame ranking, scans and analytics as a paid listing. It runs on the buyer's machine — hosting starts the day it carries a price.
SubscriptionYou set the monthly plans for your own MCPEach plan is a monthly price, the calls it includes and a rate limit, and a plan at nothing is your free tier. Trials are built in, and cancelling stops it — which is why a paid listing runs behind our gateway.
Become a publisher Read the payout terms
Earnings estimate 88 / 12 split
Paying subscribers80
Your monthly plan price$29
Gross revenue$2,320
mcprush fee (12%)−$278
Stripe processing (2.9% + 30¢ a charge)−$67
Your payout / month$1,974
≈ $23,692 a year — 85.1% of gross before the flat part. Stripe pays it out to your bank from your connected account, every Friday. The 30¢ is charged per renewal, so it is 30¢ × subscribers each month; the line above shows the percentage only, and the calculator on the publishers page subtracts the flat part in full.
Stacks

Install a whole desk, not one tool at a time

Curated bundles with a shared budget, one approval flow and a single line in your config.

All 20 stacks
Pull-request desk Product engineering

Reads the branch, runs it in a throwaway sandbox, and leaves the review on the merge request before a human opens it.

GIGISEE2NE+6 $0/mo
Bug reproduction desk Product engineering

Turns a bug report into a reproduction that fails, before anybody writes a line of the fix.

SELOPOPUFI+5 $0/mo
On-call desk Operations

Answers the page: what fired, what changed, what the runbook says, and what to tell the channel.

DAPRKUSEDI+4 $0/mo
Database change desk Data engineering

Writes the migration, proves it on a branch, and says what it will do to the table before it touches the real one.

PONEMEGI+5 $0/mo
Questions

The parts people ask about twice

A server gives the model new reach: it runs somewhere, it holds credentials, and calling it can cost money or change something outside the conversation. A skill gives the model instructions: it is text and reference files that load when the work matches, it holds no credentials of its own, and it can only use tools the model already has. Installing a skill therefore changes how your agent decides, never what it can touch. Most teams install a handful of servers, then a skill that knows how to drive them well.
88% of net revenue on the model that charges — a subscription. A free listing bills nobody, so there is nothing to split and nothing is deducted. The 88% is before Stripe's processing, which is 2.9% + 30¢ per charge and goes to Stripe and the card networks rather than to us. 85.1% of gross is what lands. The 12% covers the gateway, OAuth, call accounting, fraud, chargebacks, tax handling and the payout machinery. There is no listing fee and no minimum payout: if you earned $3.40 this week, $2.59 lands on Friday — 85.1% of it, less the 30¢ that taking one charge costs, which is why a small week keeps a smaller share of itself. You are paid through Stripe Connect — Stripe holds your bank details and the tax form, and pays out from your connected account every Friday for every sale that has cleared by then — a charge clears seven days after it is taken, in the settlement currencies Stripe supports for your country. A refund is netted off the next run rather than clawed back from your bank, and so is a dispute — with Stripe's $15 fee, which it charges the moment a cardholder's bank opens one and keeps whether the case is won or lost.
No. Your server stays where it is: it runs on your infrastructure, at an HTTPS endpoint you own, and we never receive your code and never run a container for you. What sits in front of it is the gateway — it authenticates the buyer, meters the call, applies the plan’s allowance and rate limit, records the result, and forwards the call to your endpoint with a token that says it came from us. So you are not billed for compute by us and you are not asked to package anything: a listing is an address plus a manifest. A free or open-source listing can go the other way and ship as source, which the buyer runs on their own machine over stdio — nothing to meter there, and nothing for us to switch off, which is exactly why that shape is free.
Every subscription plan carries a monthly call allowance and a rate limit, both printed on the listing before you pay. When the allowance runs out, the gateway returns a normal MCP error the agent can reason about — it never turns into a larger invoice, because the price is fixed. On top of that the account itself has a spend ceiling: it stops new paid installs rather than letting a month run away.
Publish triggers a static scan of the source and manifest, a capability-flow analysis of what each tool can reach, a secret scan, and a diff of every tool description against the previous release. Anything unresolved goes to manual review before it can be marked verified. Grades are public, including the bad ones — a C-grade server stays listed and says why. Nothing here is a promise about intent: what you approve at install is a capability budget, which is the part we can actually enforce.
No, and we would rather say so than sell a mirror nobody maintains. What the question is usually after is control over what your engineers can install, and that is a seat-plan feature rather than a second registry: seats with roles, a spend cap per seat, a key per machine that the owner can revoke, and a gate that holds a release adding a write tool until somebody on the account approves it. The catalogue stays ours and stays one thing, so a grade, a scan report and a rollback mean the same to you as to everybody else — which is the part a private mirror quietly gives up.
Claude Code, Claude Desktop, Cursor, VS Code, Zed and Windsurf install with one command; Codex, Gemini CLI, Grok, ChatGPT, Microsoft Copilot, Perplexity and DeepSeek take one entry you paste yourself, and anything built on the Agent SDK or plain HTTP connects with a key instead. The rule underneath is simple — if a client speaks MCP over Streamable HTTP or stdio, it can install from mcprush. Remote servers work in all of them, because the transport is just HTTP. Local stdio servers only work where the client can spawn a process, which rules out anything running in a browser tab. The compatibility matrix higher up this page lists what each client does with an install.
Free to browse · free to publish

Give your agent a toolbox worth trusting.

Install in one command. Pay the listing’s own price and nothing to us. Cancel, cap or roll back from the same place you installed.

Explore the catalog Become a publisher

MCP marketplace for AI agents: install MCP servers and agent skills, or sell your own

Mcprush sells two kinds of thing your AI agent can use, and both are first-class here. An MCP server gives your agent new abilities — reading your database, opening pull requests, sending a message. An agent skill is a set of written instructions that teaches it to do a job properly: a procedure to follow, the standard to hold the result to, the mistakes to avoid. You can browse servers and skills side by side, install either in one command, and pay for the paid ones on a single bill. If you make servers or skills, you can sell them here — free or by the month — and the money reaches your bank every Friday.

For customers

Search the catalogue by what a tool does, which app it works with, what it costs and what it is allowed to touch. Then copy one command. It sets up Claude Code, Claude, Cursor, VS Code, Zed or Windsurf for you — you never edit a config file by hand; Codex, Gemini CLI, Grok, ChatGPT, Microsoft Copilot, Perplexity, DeepSeek and your own code take one entry you paste yourself, and every page prints the exact shape each of them wants.

Before you install anything, you can see exactly what it will be able to do: every tool it has, what it reads, what it changes, and where it connects on the internet. You also see which version you are getting and its security grade, from a scan of that exact version. Reviews come only from people who really used it.

Most listings are free and install straight from the author’s own source, so nothing runs through us. Paid ones work through our gateway: it keeps the key, counts your calls against the plan you chose and stops at the limit, so you never get a surprise bill. Everything you subscribe to, from any author, arrives on one invoice with one card, and you set your own spending cap. Cancel whenever you like — the subscription stops at the end of the period you paid for, and anything already on your computer stays yours. In a hurry? A stack is a ready set of servers and skills that already work together for one job, installed as one thing.

For publishers

Listing is free, for a server or for a skill. For a server, point us at your repository, your npm or PyPI package, your container or your server address, say what each tool touches, and pick a price: free, or a monthly plan you write yourself — how much it costs, how many calls it includes, how fast they may come. Paid servers run behind our gateway, because that is what makes a cancelled subscription actually stop. A skill is a folder of instructions, so it is sold the same two way but once rather than monthly: one price buys the folder and every update you ship inside it, and buyers can read the whole bundle before they take it.

You keep 88% of every sale, before card fees. We are the seller on the invoice, so refunds, chargebacks and sales tax are our problem, not yours — you do not register for tax in countries you have never been to. Money reaches your bank every Friday through Stripe, which also holds your bank details and your ID check. Every version you publish is scanned before buyers get it, and the grade shown is for the version they actually install.

Listings nobody has claimed yet

Part of this catalogue was collected from public places — the official MCP registry, npm, PyPI, container registries and public repositories — so buyers can find servers that exist, not only the ones whose authors have heard of us. Those entries say not claimed on them, they credit the author’s public username, and they are always free. We do not sell anything on behalf of someone who never signed up. If one is yours, press Claim on its page and prove you own the source, and it becomes your listing. You can also ask us to correct it or take it down, and we will, without asking why.

Why a marketplace and not just a list of links

Anyone can put an MCP server on GitHub. What a list of links cannot give you is what comes after: a security scan of every release, a plain record of what each tool may read and change, one invoice instead of a dozen card charges, a spending limit you set instead of the seller, a version you can pin and roll back with the same command, and reviews from people who actually ran it.

New to MCP?

MCP (Model Context Protocol) is an open standard for connecting an AI assistant to tools and data. An MCP server is a small program that speaks it and offers a list of tools the assistant can call. An agent skill is a folder of written instructions the assistant reads when the task matches — it costs no calls, only context. Servers let an agent do things; skills tell it how to do them properly. Most teams need some of both.