Member of Technical Staff

Capy•San Francisco, CA

1d•Onsite

About The Position

Capy is building an AI software engineer that users can assign real work to. The system allows users to initiate tasks from various platforms like Slack, Linear, web, desktop, or API. Capy can then execute these tasks on cloud machines or the user's local devices, open pull requests, and manage CI and review feedback until the code is merged. The platform supports work distribution across multiple devices and cloud machines, with dedicated agents for code review and scheduled automations. The company envisions a future where developers orchestrate fleets of AI agents rather than managing single agent sessions. Capy is a small team based in San Francisco, backed by Y Combinator and prominent investors, and its product is used daily by teams for production codebases. They emphasize building Capy with Capy itself, meaning their own development processes utilize the tool. The AI coding market is competitive, and developers expect immediate results and high quality. To meet these expectations, the role requires ownership of the entire product experience and meticulous craftsmanship. "Full stack" in this context encompasses the user interfaces for directing agents, the agent loop and its associated tools, the backend for agent replanning, the underlying infrastructure, and the layer interacting with frontier AI models. This position involves making product decisions independently, designing without waiting for mockups, and solving complex engineering challenges directly.

Requirements

Agent-native. Claude Code, Codex, Cursor, etc (or your own harness) is how you already work, and you have sharp opinions about where they fall short.
Fluent in our stack. TypeScript 7 on Node 24 and React in production, and typed functional code (Effect or the like) you can read and write without a ramp.
You've shipped against MySQL, run something on Kubernetes, and touched a message bus like NATS.
Comfortable at every layer. A slow query, a broken websocket, a failing Helm release, and a misbehaving tool call in the same afternoon, and you don't wait for a specialist on any of them.
Taste in interfaces. You notice the four-pixel misalignment, you care about motion, hierarchy, and empty states, and you can point at shipped work that proves it.
Independent product judgment. You make the call on what to build and how it should feel without a PM or designer defining the problem first.
Fast without shortcuts. You verify your work, read your diffs, and would rather delete code than add an abstraction nobody uses yet.
In San Francisco. In a room with people doing the same.

Nice To Haves

Effect in production, or Rust (our remote desktop is Rust).
Sandboxed execution, VM images, remote desktops, or a git proxy you've built or operated.
A model-provider layer you've built, or direct work against Anthropic, OpenAI, and Gemini APIs.
Design-tool fluency (Figma or similar) on top of prototyping in code.
Public work: open source, posts, demos, an active account on X.
Founder or early-employee experience.

Responsibilities

Ship across the whole stack. One day is a new tool in the agent loop, the next is the React surface that shows what it did, the migration behind it, and the Kubernetes change that ships it.
Set the standard for agent UX. Threads, tasks, diffs, reviews, live machine desktops streamed from our Rust remote desktop: twenty-five agents running at once need an interface that stays legible.
Design in code. Take a feature from rough idea to a polished, animated, keyboard-first interface in the React web app and Electron desktop app.
Build the backend the agent runs on. Effect services over MySQL (Vitess on PlanetScale), realtime fan-out over NATS, the typed API and protocol every client shares, the git proxy, the worker that drains threads.
Push what the agent can do. New tools, better verification, tighter loops with CI and review, smarter task orchestration, and the provider layer across Anthropic, OpenAI, Gemini, and the rest.
Run the infrastructure. The Kubernetes clusters and IaC that deploy Capy, and the VM images and sandbox platform every cloud machine boots from.
Direct Capy on Capy. Most of your day is writing specs, steering agents, reviewing their PRs, and fixing the harness when it lets them down.
Talk to users. Developers running Capy on real codebases tell you what's wrong; you turn it into what ships that day.