Today We Fixed OAuth Discovery

And learned that sometimes the simplest solution wins

🔧 The Bug

Claude MCP couldn't find our OAuth endpoints because we hid them at:

/api/mcp/.well-known/oauth-authorization-server ❌

When Claude was looking at:

/.well-known/oauth-authorization-server ✅

One rewrite rule later... we're back in business.

ðŸŠĶ Meanwhile at Vercel...

⚰ïļ

"Here lies Claude Skills (2024-2025)"

Vercel's research revealed a painful truth about their fancy Skills system:

56%
of the time, agents simply didn't invoke the skill
79%
max success rate even when told to use skills
100%
success rate with a simple AGENTS.md file

The agent had documentation available and just... didn't use it.

Turns out making the AI decide to ask for help is harder than just... giving it the help.

📝 The Hot New Thing: AGENTS.md

Drop a markdown file in your repo. That's it. That's the whole feature.

# AGENTS.md - It just sits there

No decisions required.
No async retrieval.
No "should I invoke this skill?" deliberation.

Just vibes. And context. Always there.

Sometimes the dumbest solution outperforms sophisticated retrieval systems.

ðŸŽŊ Lessons Learned

1. Put things where they're expected (.well-known at the root)

2. Passive context beats active retrieval

3. If your agent needs to decide to use help, it probably won't

4. 40KB of bloated docs → 8KB of pure efficiency