REPO PACKAGE
better-auth agent-auth
COMMUNITY3skills
0ready to use
LABSCO SUMMARY
agent-auth-cli drives the whole flow — discovering providers, connecting an agent, managing its granted capabilities, and executing operations — through the auth-agent CLI binary; agent-auth-mcp does the identical job through MCP tools instead, for a client that talks MCP rather than shelling out. agent-auth-connectors sits on top of either one: a workflow skill for using specific connector capabilities, Gmail named as the example, once an agent is already registered and approved.
This is for a team building the server side of the protocol itself: the plugin lives in @better-auth/agent-auth for a Better Auth server, alongside a separate client SDK (@auth/agent) and CLI (@auth/agent-cli). It is not Better Auth's mainline authentication library — that ships its own separate repository and skill set — so anyone who came here expecting ordinary login-flow skills should look there instead.
READ THE FULL ANALYSIS
The protocol is not this repository's own design. The canonical specification lives at nicepkg/agent-auth-protocol on GitHub; this repository is Better Auth's implementation of it, demonstrated in Next.js reference apps under examples/ (paired with Drizzle), two of which add WebAuthn/passkey sign-in on top of the baseline email/password flow.
Two of the three skills are the same job through two different doors. A reader picks agent-auth-cli or agent-auth-mcp depending on whether their own client shells out to a CLI or speaks MCP, then reaches for agent-auth-connectors only once that choice is working.
Checked 19 September 2026 from the repository README and its three currently live skills.
WHAT'S INSIDE
3 showing · 3 total3 of 3
SETUP
Showing all 3 skills
agent-auth-cli
An agent cannot be trusted with your password, so it gets its own named, approved permissions instead — and this does all of that from the terminal.
agent-auth-connectors
The manners and the pitfalls of letting an agent into someone's email account — connect once, grant only what the job needs, and warn the person before you do.
agent-auth-mcp
Inside Cursor or Claude Desktop, the assistant asks for permission itself — it requests only the actions a task needs, waits for you to approve, and then runs them.
No skill in this package matches that.
HOW TO GET IT
WHOLE PACKAGEInstall the whole package (3 skills)
npx skills add better-auth/agent-authSINGLE SKILLOr install a single skill
npx skills add better-auth/agent-auth --skill <name> --full-depthPick the skill name from the Skills tab — each entry there installs independently.