We were drowning in AI coding agents. So we built a local control plane.

I did not plan to build an observability product. I planned to ship features with AI coding agents. Then the agents multiplied. Cursor in the IDE. Claude Code in one terminal. Codex in another. OpenCode for free-model experiments. Each one was useful. Together they created a new kind of mess: spend you cannot see, sessions you cannot search, and zero shared picture of what "the agents" did yesterday.

That mess is why superbased exists.

The real pain was not autocomplete

Autocomplete is solved enough. The pain showed up when agents started acting like junior teammates:

• Long sessions that quietly burn tokens

• Parallel tools with different pricing and logs

• No single place to answer "what did this cost?" without opening four UIs

Cloud AI analytics vendors want you to upload traces. That is a non-starter when the trace is your private repo and half-finished product thinking. So the bet became:

• Local-first control plane.

• One observer on your machine.

• Many agent adapters.

• Dashboard on localhost.

• Optional proxy when you want provider-exact counts.

What we shipped SuperBased (`@superbased/observer`) is open source and installs like this

npm install -g @superbased/observer

observer start

You get a local dashboard, cost rollups, and adapters across the tools people actually use. No account required for the core loop. We also learned the hard way that SEO still thinks we are an old screenshot app. Fixing listings and writing real demos (OpenCode + Groq, then Medium/Substack/Dev.to) has been as much of the job as the binary.

What I would tell another founder

1. If your users run multiple AI tools, "usage page inside one vendor" is not a strategy.

2. Privacy is not a slogan when the data is agent transcripts. Local-first is a wedge.

3. Content works better when each platform gets an original angle, not the same post five times.

Links

https://superbased.app

Dev.to thesis: https://dev.to/superbased/one-control-plane-for-claude-code-cursor-codex-and-opencode-3bmc

Substack (OpenCode cost walkthrough): https://superbased.substack.com

Happy to answer questions about adapters, local SQLite capture, or how we think about proxy vs log-based costing.