jaredrhod/fullstack-agent: An Installer Wizard That Assembles Memory, Voice and Face Around Claude Code

The problem is assembly, not intelligence

Most people asking for an "AI agent" already have a model. What they do not have is persistence, a voice channel, a face, and a way to move things around a screen. fullstack-agent takes four separately maintained repositories and installs them as one agent, in a single conversation, on your machine. The README is blunt about the framing: "Not an agent that writes full-stack code. An agent that HAS a full stack." That distinction matters, because anyone arriving from the phrase "full stack agent AI" may expect code generation and get a personal assistant instead. The audience is narrower than the name suggests. You need Claude Code and a Claude subscription, and the README states the $20 Pro plan is enough. The wizard is not a shell script you configure; it is a Claude Code session that asks questions. If you do not already live in Claude Code, this is the wrong entry point, and the README points such readers at jaredrhod.com first.

Four repos, one conversation, and a wizard that adopts before it installs

The architecture is assembly, not invention. Each piece is its own repository: ai-memory-vault for persistent memory built on plain text files, backtalk for hold-a-key voice conversation, ai-visualizer for full-screen visualizers that react to idle, listening, thinking and speaking states, and barehands for webcam hand tracking that opens in its own window rather than replacing the face. The wizard asks which pieces you want and explains each in plain English before you decide. Wiring happens through configuration paths rather than by moving your data, which is why the README can say your vault stays wherever it already lives. The interesting design choice is the adoption pass. If you previously hand-built a memory vault, voice lines or a visualizer, the wizard detects them first: agent identity and vault are kept and never rebuilt, hand-built voice and visualizer code is replaced in place with the maintained repo version while the old build stays on disk, and a custom visualizer scene is copied into the gallery as your own face. The stated reason for replacing hand-built voice and visualizer code is that the repos carry a year of fixes and improve with a git pull, while a hand-built version is frozen. That is a defensible trade, though it does mean the wizard decides which of your work counts as identity and which counts as an implementation detail.

Installing fullstack-agent on macOS, Linux and Windows

The README gives one paste per platform. On Mac and Linux you create a folder in your home directory, clone the repository, and start Claude Code with the prompt "set me up".

mkdir -p ~/my-agent && cd ~/my-agent && git clone https://github.com/jaredrhod/fullstack-agent && cd fullstack-agent && claude "set me up"

Claude Code opens with the installer already talking to you. It asks for your agent's name and personality, offers the author's own Jarvis preset, asks which pieces you want, and asks where your notes live. The Windows path is a single PowerShell block that downloads the repository as a zip instead of cloning, so it works on a machine without git, then launches Claude Code if it is on the path.

$d="$env:USERPROFILE\.local\bin"; if (Test-Path "$d\claude.exe") { $env:Path="$d;$env:Path" }; New-Item -ItemType Directory -Force -Path $HOME\my-agent | Out-Null; cd $HOME\my-agent; if (-not (Test-Path fullstack-agent\fullstack-agent.md)) { Invoke-WebRequest https://github.com/jaredrhod/fullstack-agent/archive/refs/heads/main.zip -OutFile fsa.zip; Expand-Archive fsa.zip . -Force; New-Item -ItemType Directory -Force -Path fullstack-agent | Out-Null; Get-ChildItem fullstack-agent-main -Force | Copy-Item -Destination fullstack-agent -Recurse -Force; Remove-Item fullstack-agent-main -Recurse -Force; Remove-Item fsa.zip }; cd fullstack-agent; if (Get-Command claude -ErrorAction SilentlyContinue) { claude "set me up" } else { Write-Output "Claude Code is not installed yet. Install it first at https://jaredrhod.com/start then paste this again." }

Two details from the README are worth reading before you paste. The block is safe to run repeatedly: it skips the download when fullstack-agent.md is already present, and if an earlier attempt died partway it downloads again rather than assuming the job was done. And the Claude Code installer on Windows downloads about 330 MB while printing nothing, so leave that window alone until it reports installation complete. After setup you get three Desktop shortcuts named after your agent: Chat for a typed terminal session, Talk for voice and face, and Barehands for voice plus the hands board. They run start.sh, or start.bat on Windows, if you prefer the terminal. Ctrl-C in the window stops them.

Claude Code only, and the wizard is the installer

The hard constraint is the platform. fullstack-agent runs on Claude Code only, and the installer itself is a Claude Code wizard. There is no standalone setup binary you can run in CI, no headless flag, and no documented non-interactive path; the README's install instructions all end in a claude invocation with a natural-language prompt. If your team provisions machines from a configuration management tool, this repo does not fit that workflow, and the conversational steps (agent name, personality, which pieces, where notes live) would need to be answered by a human or scripted around the wizard entirely. A second limitation is the folder location. The README says the agent lives in a folder directly in your home directory on purpose, because on Macs things that run in the background out of Documents get silently blocked by the system. That is a real macOS behaviour, but it also means the stack is tied to a user home directory rather than a project checkout, which complicates shared or service-account deployments. Finally, the README does not document rollback. It states the wizard never deletes, overwrites or moves anything you built, and that replacements retire the old thing in place, but there is no described command to undo an install and return the machine to its prior state.

What it is not, and what to compare it against

Do not compare this to agent frameworks. LangGraph, CrewAI and similar projects give you primitives for building agent control flow in your own code; fullstack-agent gives you a finished personal setup, installed and wired, with no programming surface described in the README. The closer comparison is a hand-rolled Claude Code setup: a CLAUDE.md, a notes folder, and whatever voice or visualizer scripts you wrote yourself. The difference in approach is maintenance. A hand-built stack is frozen the day it was written, as the README puts it, while the four underlying repos keep improving and can be pulled forward. That is the actual argument for adopting the wizard rather than copying the pieces manually, and it is also the argument against it: you are accepting four upstream repositories and their update cadence in exchange for not maintaining the glue yourself. If you want one component only, say voice, installing backtalk directly is a smaller commitment than running a wizard that will also offer you memory, a face and hands.

Updates, licence and what maintenance actually costs

Updating is documented as one command on macOS: ./fullstack-agent/update.sh. On Windows the README tells you to ask your agent to "update everything and tell me what changed" rather than naming a script, even though update.bat exists in the repository listing. The reassuring part is the data boundary: your files live outside the repos, so the README states updates never touch who your agent is or what it remembers. The recurring cost is not the update command, it is that you now depend on four repositories plus Claude Code itself. The repository is licensed AGPL-3.0, which is a strong copyleft licence; if you plan to modify this and offer it to others over a network, read the licence text and, where it matters, take proper advice, because the obligations of AGPL-3.0 are broader than those of MIT or Apache-2.0 and this article is not legal advice. The last push to the repository was on 2026-08-30, and the repository is not archived.

Editorial conclusion

Adopt it if you already pay for Claude Code and want the four jaredrhod pieces wired together without hand-assembling them, especially if you previously built a memory vault or visualizer by hand, since the wizard adopts identity and vault and promotes a custom scene instead of overwriting it. Skip it if you want a code-generating agent, work outside Claude Code, or need a non-interactive installer for fleet provisioning: the whole flow is a conversation, and the README states the wizard itself is a Claude Code wizard. Verify first that your vault location is the one you intend to keep, since the README says the vault stays where it lives and pieces connect by configuration paths, and confirm on your own machine that the Windows PowerShell block finishes before you judge the setup, because that path downloads the toolbox as a zip and the Claude Code installer prints nothing while downloading roughly 330 MB.

Frequently asked questions

Official sources

1. Issues
2. jaredrhod/fullstack-agent on GitHub
3. License: AGPL-3.0
4. README