Skills have become a common building block for agentic workflows. Most (if not all?) coding agent harnesses support them. I treat them very much as dependencies within the system.

Managing agent skills across tools and repositories is a copy-paste nightmare. We've built a tool called dotagents that treats them as versioned, lock-filed dependencies they actually are.

dotagents will:

• Help you manage versions and update skills, treating them like dependencies
• Deduplicate configuration for skills/mcp/hooks across agent harnesses
• Define trusted sources to reduce a risk of pulling random skills from untrusted sources.

You can start using dotagents by either installing it globally:

npm install -g @sentry/dotagents

or using it over npx. Set up by running

npx @sentry/dotagents init

It will take you through an interactive wizard that will help set up your initial agents.toml file. It will look something like this:

version = 1
gitignore = false
agents = ["claude", "cursor"]

[trust]
github_orgs = [ "anthropics" ]
github_repos = [ "getsentry/skills" ]

[[skills]]
name = "code-simplifier"
source = "getsentry/skills"

Why use dotagents?

Dependency Management

Oftentimes skills are useful across multiple repositories in your organization and the current pattern is to pack them up in a common repository and either install them directly or copy-paste them to a repository. It's hard to keep up to date - you have to do that mostly manually - and difficult to deduplicate that across multiple coding agents (.agents vs .claude etc.).

There exist tools like npx skills which help with skills discovery and installation. However, what I wanted here is something that can help me manage more than just skills (MCPs mostly) and treat the skills as dependencies with version pinning and defining trusted sources.

Security challenges

The main reason to hesitate to use random skills is that they are a new (and very scary) supply-chain attack vector.

You can read about a great example of a really bad attack here. In this case, the attacker injected a malicious skill into one of the skill marketplaces called Clawdhub (meant for OpenClaw assistant agents). The skill simply injects a malicious prompt into one of the skill files without even trying to disguise it.

Even if you read the skills you are adding to your arsenal, you can very very easily miss any malicious instructions. They could be in a file referenced by main SKILL.md (as is described in the article above), could be specifically hidden inside an HTML comment - which will cause it to not render in the pretty GitHub reader. They could also just recommend the agent to install a package that looks innocent but can then be used to exfiltrate data from your machine (i.e install react-super-awesomeinator to validate all react changes).

The point is - it's very hard to catch. The attackers have a new vector that can get you pwned in no time. This is even harder if you use any type of skills search engines, where people can maliciously post skills targeted at pwning you.

dotagents allows you to define trusted sources:

version = 1
agents = ["claude", "cursor"]

[trust]
github_orgs = [ "anthropics" ]
github_repos = [ "getsentry/skills" ]

In this example, if a developer tries to add a skill from other sources, they should be stopped and limited to only specific trusted repositories. Using the example of Sentry's skills we could white-list the usage of those and then make sure that we run additional safety checks/skill scanner, or external tool to help us build conviction and trust.

MCP and hooks

We have experimental and limited support for MCP and Hooks management. MCP support should work across our set of supported agents and inject MCP configurations to the respective spots in the configuration files. The same is true of hooks, but as not all agents support hooks (or they diverge too much), the utility is a bit more limited for now. Here's an example of defining MCP server with dotagents:

[[mcp]]
name = "sentry"
url = "https://mcp.sentry.dev/mcp"

Feedback welcome

dotagents is still early and we'd love to hear what you think. Tag me on Twitter/X or open an issue on the repo.