fullstack-agent: a Claude Code wizard that assembles four jaredrhod repos into one agent

The problem is assembly, not code generation

The README is explicit about what this is not: "Not an agent that writes full-stack code." The full stack in the name refers to the agent's own faculties, described as memory, voice, face and an optional set of hands. Each of those already exists as a separate open repository under the same author: ai-memory-vault for persistence, backtalk for push-to-talk voice, ai-visualizer for full-screen animated faces, and barehands for webcam-driven manipulation of notes and images on screen.

The work this repo removes is the wiring between them. Four repositories, each with its own configuration, its own paths and its own assumptions about where your notes live, become one guided conversation. The README frames the audience as people who have never used Claude Code and who want a working agent rather than a pile of cloned folders. It also names a second audience that is easy to miss: users who already hand-built some of this from the author's older prompts, and who are told the wizard "adopts before it installs."

A wizard, not a daemon: what runs where

The primary language of the repository is Batchfile, which tells you something about its scope. There is no long-running service here. The installer is a Claude Code wizard, meaning the orchestration logic is a conversation the model conducts, and the Windows entry point is a PowerShell one-liner that downloads a zip and hands control to `claude "set me up"`.

The runtime lives in the four constituent repos. After setup, the three Desktop shortcuts named after your agent are described as thin wrappers: they "just run fullstack-agent/start.sh, or start.bat on Windows." So the data flow at runtime is start script, then whichever combination of memory, voice, visualizer or hands you selected. The installer's job is to produce that start script plus the configuration paths that connect the pieces.

One design note is stated plainly and is worth taking at face value: the agent folder sits directly in your home directory on purpose, because on Macs background processes launched from Documents get silently blocked by the system. That is a real macOS restriction, and it explains a layout choice that would otherwise look arbitrary.

The install commands and the Windows zip workaround

Mac and Linux get a three-part shell line: create `~/my-agent`, clone the repository into it, then run `claude "set me up"` from inside the cloned folder.

Windows gets a much longer PowerShell block. It prepends `$env:USERPROFILE\.local\bin` to the path if `claude.exe` is present there, creates `$HOME\my-agent`, and then, instead of using git, downloads `https://github.com/jaredrhod/fullstack-agent/archive/refs/heads/main.zip`, expands it, and copies `fullstack-agent-main` into a `fullstack-agent` directory. The README states the reason: the zip path "works on a machine with no git installed," and the installer then sets up git during setup. The same block is written to be re-runnable. It skips the download when `fullstack-agent\fullstack-agent.md` already exists, and if a previous attempt left a partial folder it downloads again rather than assuming success.

Two operational details are called out. The Claude Code installer on Windows pulls roughly 330 MB and prints nothing while it does, so the README warns you to leave the window alone until it reports Installation complete. And the requirements are narrow: Claude Code with a paid subscription, with the $20 Pro plan described as sufficient.

Adoption of existing hand-built pieces, and what gets replaced

This is the most specific behaviour in the README and the part most likely to matter to returning users. The wizard distinguishes three cases.

Your agent identity and your vault are found and kept. The README says they are "never rebuilt, never moved," and that the wizard will not re-ask questions you already answered. Hand-built voice lines and visualizers are replaced, on the stated grounds that the repositories carry a year of fixes and keep improving with a `git pull`, while a hand-built version is frozen. The old build stays on disk. Custom visualizer scenes are the exception in the other direction: if your AI previously built you a scene, the wizard copies it into the visualizer's gallery as your own face.

The general claim is that the wizard "never deletes, overwrites, or moves anything you built," and that replacements retire the old thing in place. That is a promise about installer behaviour, and it is the kind of promise worth confirming against your own backups rather than accepting from a README.

Platform and model constraints

The hard limit is stated in the first lines: this runs on Claude Code only, and the installer is itself a Claude Code wizard. There is no documented path to another model provider or to a plain shell install that skips the conversation. If you are standardized on a different assistant, none of the four components is reachable through this repository's documented route.

The second constraint is the platform split. macOS and Linux get a git clone and a shell script. Windows gets a zip download, a copy step, and later a `start.bat` and an agent-mediated update command rather than a script. That asymmetry means Windows users have one fewer scripted entry point, and the update path there is a natural-language request to the agent rather than a single command.

A third limitation is implied by the troubleshooting model. The README tells you that when something breaks, you should ask your agent to fix it, because each repo ships a troubleshooting guide written for the agent to read and the agent is instructed during setup to do the fixing. That works when the failure is inside the stack. It does not help when the failure is Claude Code itself, or when the agent cannot start, which is exactly the state in which you would need the guide.

Compared with wiring the four repositories yourself

The obvious alternative is to clone ai-memory-vault, backtalk, ai-visualizer and barehands individually and connect them by hand. The difference is not capability. It is who holds the integration knowledge.

Doing it manually means you own the configuration paths between the pieces, you decide where the vault lives, and you write your own start scripts. You get the same four repositories and you can read every line before it runs. The cost is that you reproduce the wiring the wizard performs, and you inherit the same platform quirks, including the macOS background-process restriction that pushes the working folder into your home directory.

Using the wizard means an LLM-driven conversation performs that wiring, with the adoption rules above applied automatically. You trade direct control of the integration step for not having to understand four separate configuration surfaces. For a user who has never opened Claude Code, that trade is the entire point of the repository. For a user who already maintains hand-built versions of two or three of these pieces, the adoption logic is the deciding feature, and the replacement of hand-built voice and visualizer code is the part to think about before running it.

Updates, licensing and maintenance cost

Updating is asymmetric across platforms. On macOS the README gives `./fullstack-agent/update.sh`. On Windows there is no scripted equivalent in the material: you ask the agent to "update everything and tell me what changed." In both cases the claim is that your files live outside the repos, so updates do not touch agent identity or memory. That separation is what makes a `git pull` safe to run, and it depends on the vault staying where it already is, connected "by configuration paths, not by relocation."

The licence is AGPL-3.0. That is a copyleft licence with a network-use clause, and it applies to this installer repository. The four component repositories may carry their own terms, which the README does not state, so check each one before you build anything you intend to distribute or host for others. Nothing here is legal advice.

Maintenance cost is mostly the agent's, not yours, by design. The repos keep improving with a `git pull`, and the wizard is written to be re-run to add barehands later. The recurring cost you actually pay is a Claude subscription and the habit of opening Claude Code in the agent's folder, which the README calls the daily routine.

Editorial conclusion

Adopt it if you already pay for Claude Code, you are comfortable with the agent living in a folder in your home directory, and you want memory, voice and a visualizer wired together by a conversation rather than by hand. Do not adopt it if you need a non-Claude model, a headless server deployment, or a Windows setup you can fully script without the wizard. Before you commit, verify three things on your own machine: that the Windows PowerShell paste completes and leaves fullstack-agent.md in place, that start.sh or start.bat launches the Chat shortcut, and that update.sh on macOS leaves your vault and agent identity untouched.

Official sources

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