Five problems I couldn't buy my way out of.

No agent tooling could be adopted that was unable to say who did what. So I designed the governance as the substrate rather than a wrapper: every action executes under the requesting human’s identity, passes a tiered authorization gate, and is written to an append-only, hash-chained ledger.

Designed this way because nothing off the shelf survived the compliance review.

Agent browser tools hand the model a DOM snapshot and hand the human a log file, so neither sees what the other did. Brainbow streams a real Chromium at 30fps to a viewer while the agent drives that same session over MCP or REST.

Built because debugging a headless agent by reading transcripts is miserable.

Every agent platform eventually stalls waiting on someone to write the next connector. Synth lets the agent synthesize the tool it needs at runtime — and refuses to execute any of it without an explicit human approval, recorded.

Built because the tool registry was always the thing holding the work up.

In most studios one person knows why that cable goes there, and when they are out the session stops. PatchBay makes the room itself the record — real gear, real jacks, real signal paths, captured once and readable by everyone.

Built for musicians, not engineers. If a patch is hard to explain, the patch is wrong.

Plenty of teams cannot put an agent runtime behind someone else’s login. This is mine, MIT licensed and self-hosted end to end. Building it in the open forced every enterprise assumption out of the core and into configuration.

Built because "trust us with your infrastructure" is not an answer for everyone.