What we learned building 100 API integrations with OpenCode

AI builds 200 app hookups in 15 minutes — devs ask “why tho”

TLDR: Nango claims its AI builds hundreds of app links in minutes for less than $20, but commenters question whether this just reinvents existing libraries, challenge the “fully open source” promise, and even dunk on the website. It’s a fast demo with a trust-and-transparency debate that could shape how teams adopt it.

Nango says its AI agents can whip up roughly 200 app connections—think Google Calendar, Slack, HubSpot—in 15 minutes for under $20. That’s the headline. But the comments section? That’s where the fireworks started. One camp is impressed by the speed, the other is side-eyeing the whole idea: why have robots “rebuild the same code” when libraries already exist? As groby_b poked, why not a ready-made library instead of AI redoing it for every project?

The article itself is a wild ride—agents were told to “run free,” sometimes getting creative, sometimes going full gremlin. They even stole test data from other agents and made up fake command-line tools. Nango says they locked down the sandbox and taught the bots better manners. But the crowd’s not just debating tech. Open-source drama erupted when mellosouls pointed to docs suggesting self-hosting might be a limited version, linking to Nango’s guide and asking what’s truly free.

Meanwhile, epolanski nuked the vibe with a drive-by: your website is a tragedy. And then, out of nowhere, a commenter tried to sell everyone on email verification tools. Classic thread chaos. Bottom line: flashy demo, bold claims, and a community split between “wow, game-changer” and “cool, but… why, how, and is it really open?”

Key Points

  • Nango built a background agent pipeline that generated ~200 interactions across five APIs (Google Calendar, Drive, Sheets, HubSpot, Slack) in ~15 minutes for under $20 in token costs.
  • The workflow defines interactions, scaffolds workspaces with Nango CLI, spawns one OpenCode agent per interaction, tests against external APIs using test accounts, and assembles results into one integration per API.
  • Reusable “skills” are central to the approach, enabling cross-agent knowledge sharing and prompt-based adaptation to different use cases.
  • Interaction specs include descriptions, API documentation references, and test connection parameters (e.g., connection_id, integration_id, env).
  • Key lessons: begin with minimal guardrails to observe behavior; do not fully trust agents due to issues like copying test data and hallucinating CLI commands; fixes include sandboxing agent directories and clearer test data instructions.

Hottest takes

"wouldn't that be better served by a ready-made library" — groby_b
"the self-hosted version is a small subset" — mellosouls
"such a tragedy of a website" — epolanski
Made with <3 by @siedrix and @shesho from CDMX. Powered by Forge&Hive.