We build local-first developer tools, Google Cloud foundations as code, and AI agents that open pull requests under their own names. All of it ships from one public monorepo, through one merge queue.
# Wren takes its own identity, not a human's $ eval "$(bazel run //tools/agent-app -- env wren)" $ gh pr create --fill ✓ opened by vitruvian-wren-agent[bot] # A different agent reviews it ✓ approved by vitruvian-aegis-agent[bot] $ bazel run //tools/watch-pr ● bazel test //... passed ● merge queue landed ● post-merge pipeline running $
Each one started as something we needed. They are maintained in the monorepo and published to their own repositories.
A local development orchestrator. One CLI provisions container VMs, ephemeral databases and cloud emulators, local K3s clusters and Cloudflare tunnel ingress, and starts everything in dependency order.
brew install VitruvianSoftware/tap/devx
Declarative multi-node K3s clusters on macOS. Describe the nodes in one file and homelab provisions Lima VZ virtual machines, networking and the cluster to match.
brew install VitruvianSoftware/tap/homelab
Reusable Pulumi ComponentResources for Google Cloud, published as Go and TypeScript packages: projects, networks, IAM, logging and the rest of an enterprise landing zone.
A Slack MCP server with dual-token auth: around forty tools covering channels, threads, search, files, user groups and Canvas editing, acting as the bot or as you.
A macOS menu bar app and Telegram bot that forwards your messages to a locally installed AI coding CLI, so a full agent session is reachable from your phone.
Sign in with an OAuth provider and see exactly what came back: decoded tokens, claims, scopes and user info. React and Express in one container on Cloud Run.
Every change to our cloud systems takes this route, and agents have no other.
Every change starts on its own branch in its own git worktree, so parallel sessions never share a checkout.
One Bazel graph builds and tests Go, TypeScript, Python, Kotlin and Swift, so a laptop and CI agree.
A second identity approves. Agents review each other as GitHub Apps; nobody approves their own work.
The queue tests each change against the latest main before it lands. Agents never merge around it.
Pipelines deploy with short-lived credentials, and Copybara exports each project to its public repository.
Two thousand years later it is still a good test for software.
Reproducible builds, everything as code, and a guard written for every incident so the same failure cannot happen twice.
We build what removes friction from our own work first: one command to bring up a stack, one to tear it down.
Clear interfaces, honest documentation, and one design language shared across our apps, from this site to the Android remote.
Ten coding agents work in our monorepo. Until recently, GitHub saw all of them as one person, so none of them could review another's work. Here is how we gave each agent its own GitHub App, what we had to...
ReadThe era of “chat wrappers” is completely dead. As of Q2 2026, the transition from basic text-generation LLMs to fully autonomous agentic code workflows is accelerating rapidly. However, a stark divide is emerging between teams trying to prompt their way...
ReadMeet ADK Agents - the intelligent DevOps assistant that understands your codebase, executes commands safely, and helps you ship better software faster. Built on Google's ADK framework with developer productivity at its core.
ReadEvery tool, pipeline, guardrail and incident fix described on this site is in vitruvian-core, in the open.