Article
We built Agent Marketplace to share AI agent resources across coding tools, and we're releasing it as open source. Here's what it does and how to get started.
AI coding tools are now part of how our engineers work every day. But each tool keeps its setup in its own folders, file formats, and configuration fields, so the instructions you write for one tool rarely work in another.
Today we're releasing Agent Marketplace, a command-line tool we built at Adyen to share AI agent resources across teams, repositories, and tools. Agent Marketplace is an open source project built by Adyen engineers. While this isn't an Adyen product or service, it’s released under the Apache 2.0 license for anyone to use, and we’re excited to share this with the wider developer community.
What Agent Marketplace does for you #
You pick the resources you need. The tool fetches them, translates them into the format each of your AI tools expects, and writes them where that tool looks for them.
One configuration file stays the source of truth. You can delete every generated file and rebuild it with a single command, so a new device takes minutes to set up rather than an afternoon.
For platform engineering teams, the groups that run shared developer tooling inside a company, publish a resource once instead of writing the same setup instructions for every tool. For everyone else, the tool lets you search for what you need from the terminal, in an interactive browser or with a search command, and keep it current with one command. Nobody has to learn repository layouts, YAML schemas, or tool-specific configuration formats to get started.
What you can share #
Agent Marketplace handles five kinds of resource:
- Skills give an agent instructions and reference files for a specific task.
- Commands are named workflows that people run directly.
- Subagents are agent profiles with their own instructions, tools, and model settings.
- Hooks are scripts that run on events, such as submitting a prompt or calling a tool.
- MCP servers connect agents to outside tools through the Model Context Protocol (MCP), the open standard weadopted last year .
On top of those sits a collection, a named bundle of the five kinds that can draw from several registries at once. A collection installs nothing of its own. Turning it on turns on its members, so a team can publish one bundle for a role and a new joiner installs the lot in one go, while still opting out of any single member.
Resources come from a Git repository, a local folder, or a hosted Marketplace that indexes registries so you can search them in one place.
How it works #
Your configuration file lists the resources you want and the tools you use. Run agent-marketplace sync, and the tool reads from a Git repository or a marketplace service, checks the files against a pinned version, and writes them into place.
Running sync twice does nothing the second time, so it's safe in a script or a CI pipeline. It updates the files it manages and leaves everything else alone. If a name clashes with something you wrote yourself, the tool reports it rather than overwriting your work.
Version updates stay separate from syncing on purpose. A sync installs the versions you pinned, so your setup doesn't shift under you between runs. When you want newer versions, run update and move the pins deliberately.
Settings merge across four layers: your organization, your machine, the repository you're in, and your own local overrides. Platform engineering teams set the baseline centrally, project owners add the context for their codebase, and individual developers keep their personal preferences. Each group controls its own scope without stepping on the others.
Write a resource once and it compiles for eight tools: Claude Code, Codex, Cursor, Devin CLI, Factory Droid, Gemini CLI, Opencode, and Windsurf. The AI landscape moves quickly, and locking your team's workflows to one vendor's format is a cost you pay later. Where a target tool can't support a setting, the compiler says so instead of writing a file that fails quietly.
Why we built it #
Over a thousand developers at Adyen use AI tools in their daily work. Our engineers stay in the loop on what these tools produce, and people review the output, as we describe in AI in our development lifecycle. Sharing usually starts informally. Someone writes a good set of instructions, posts it in a channel, and other people copy it into their own repositories. That works for a while. Then copies multiply, owners are unclear, nothing is versioned, and a resource written for one tool doesn't work in another. The sharing itself was the win, so we kept it and added structure: one place to publish and find resources, with versions and content checks that make an install reproducible.
How resources are checked, and what stays with you #
Agent Marketplace checks resources at three points.
- Registries . In a hosted Marketplace, the operator (the person or team who runs it) decides which registries are approved and sets license, size, and hidden-content policy. The operator can revoke a whole registry, every revision of one resource, or one exact revision.
- Before publishing . The Marketplace scans each resource and its files for things such as hidden text, invisible Unicode characters, executable scripts, remote URLs, tool grants, and hook or MCP capabilities. Each finding is rejected, flagged as a warning, or allowed. A rejected revision isn't served.
- On your machine . The tool installs exact revisions and verifies the digest of what it downloads. If a revision has security findings, it shows them and asks you to confirm before anything is written. Revoked revisions are blocked, andsync --check reports them.
These checks reduce risk. They don't prove a resource is safe to run. Hooks and MCP servers can run scripts and open network connections, so read what a resource does before you install it. If you run a Marketplace, you also decide which registries to approve and who can reach it.
What data the tool collects #
Agent Marketplace doesn't collect data about you or how you use it. Telemetry is off by default, and this release has no project endpoint for data to go to.
If you want usage signals as a platform engineering team, turn telemetry on and send it to an endpoint you own. At the basic level, that's a count and duration for each command, with the command name, the target tools, and whether it succeeded. At the detailed level, it also includes a snapshot of your configuration, including registry and resource names, so treat that data as internal to your organization. The tool never sends prompts, search text, command arguments, usernames, or resource contents.
A tool for distributing developer resources shouldn't turn into a way to watch developers.
Getting started #
Agent Marketplace ships as a single binary for macOS and Linux with no runtime dependencies.
Search a registry, install what you need, and sync:
agent-marketplace init # choose the AI tools you use
agent-marketplace install --from <registry-url> # take everything in a registry
agent-marketplace sync # write it into your tools' folders
agent-marketplace status # see what landed, and where it came from
The same setup works on a personal machine. A short configuration file and one command keep your own environment easy to copy and easy to clean up, which is why several of us wanted this at home as well as at work.
Hooks are experimental in this release. They compile and write files from end to end, but event handling and data formats still differ between tools. Check the generated files before you rely on them in production.
Read the install guide and documentation, or view the source on GitHub.
Agent Marketplace is maintained by the Agent Marketplace team at Adyen. If you spot a bug or want to request a feature, open an issue on GitHub.
We're looking forward to seeing what you share with it.