Rob Pruzan spent the better part of several years at Vercel hunting down React performance bottlenecks—the kind of work that eventually pushed his open-source diagnostic tool, React Scan, past 22,000 GitHub stars. But somewhere along the way, his attention drifted toward a messier problem: how do you actually manage a half-dozen AI coding agents working simultaneously without the whole thing descending into chaos?
His answer is Zenbu, an IDE that treats agent wrangling not as a feature tacked onto a traditional code editor, but as the core design challenge. The YC-backed startup emerged earlier this year with a pitch that sounds almost counterintuitive—this is an editor you're supposed to tear apart and rebuild while you work.
An editor designed for agents, not retrofitted
Most developer tools bolted AI capabilities onto existing frameworks. Zenbu reversed the equation. According to the company's Y Combinator profile, it bills itself as "the extensible IDE for orchestrating coding agents"—software conceived from scratch around workflows where multiple AI agents grind away in parallel on different corners of a codebase.
The central bet: managing several agents at once demands different building blocks than chat windows and autocomplete suggestions. Zenbu's features reflect that conviction. Split panes let you watch multiple agents side-by-side. Git worktrees isolate each agent on separate branches, sidestepping merge conflicts. Built-in PR creation ships their output without manual shuffling between tools.
The IDE also integrates Pi, a terminal-based coding agent with its own extension ecosystem, letting developers fire off Pi commands and tap multi-provider AI backends without leaving the interface.
But the truly unusual part happens the moment you install it.
Zenbu delivers its own TypeScript source code to your machine—raw and uncompiled—dropping into a ~/.zenbu directory tracked by Git. The application dynamically compiles and executes that code inside Electron, with hot reload running in both the main process and renderer. Tweak a plugin, watch it update live. Fork the behavior. Version your modifications. The whole thing is meant to be pulled apart.
The framework beneath
That design stems from Zenbu.js, an MIT-licensed JavaScript framework Pruzan built specifically for what he terms "hackable software." The framework's documentation describes a plugin injection model with "advice" mechanisms—code that intercepts and reshapes application behavior on the fly. Services, database hooks, RPC layers, event systems. All exposed for runtime tinkering.
The README doesn't bury the philosophy: production apps should be as malleable as development builds. Zenbu.js reached version 0.4.3 on June 4, suggesting Pruzan was iterating rapidly in the days surrounding the product's public unveiling.
A technical review published by developer Florian Narr in May flagged the alpha-stage maturity and testing coverage gaps, while noting the audacity of shipping applications as editable source. The current build runs on Electron for macOS. Windows, Linux, and alternative runtimes like Tauri remain listed as roadmap items—in progress, not yet delivered.
Arriving into a crowded field
Zenbu lands at a moment when the market is already congested. Established players have been layering in agent management over the past year or so—Cursor introduced a web interface for dispatching background agents back in mid-2025, then added trigger-based automations a few months later. Google positioned its Antigravity IDE around agent workflows at its developer conference last year. A cluster of smaller entrants—Invoker, Capella, Barnaby, Stagewise—surfaced throughout the year with overlapping promises.
What sets Zenbu apart, at least in theory, is the wager on radical customizability. Where most IDEs offer plugin APIs with safety rails firmly in place, Zenbu hands you the source files and says change whatever needs changing.
Whether that's liberating or just introduces another layer of instability depends largely on your appetite for wobbly foundations.
The company is a solo operation—Y Combinator's listing shows a team of one—running on what appears to be standard accelerator deal terms. No external funding rounds or pricing structures have been made public. The macOS download is available directly from zenbu.dev with no obvious paywall.
Unanswered questions
Practical details remain murky. The site mentions "multi-provider support" through Pi, but doesn't enumerate which AI backends actually work or how smoothly they interoperate. Performance specs for running multiple agents in parallel worktrees aren't documented anywhere public. And the long-term business model—whether this stays open-source infrastructure or morphs into a commercial offering—hasn't been spelled out.
For developers already cycling through Cursor tabs and terminal windows while trying to herd agent output into something coherent, Zenbu proposes a different frame: make the orchestration itself the product, then make every piece of that product editable.
Whether that degree of control turns out to be essential, or just adds fresh complexity to an already complicated workflow, probably depends on how fast coding agents evolve from helpful assistants into genuinely autonomous collaborators.
The framework lives on GitHub. The IDE runs on macOS, for now.
The rest is still compiling.
