Tech • AI • Robotics • Game
Is Claude Fable 5.1 Actually AGI? I Tested It Inside My JARVIS
7/10TL;DR
A voice-controlled Jarvis setup built on Claude Code, local memory tools, and connected business apps is being used to automate email triage, CRM follow-up, and invoicing with spoken commands.
KEY POINTS
How the system is structured
The setup uses a desktop Jarvis app as the voice and interface layer, while Claude Code runs underneath through an agent SDK. The assistant is connected to workspace tools, browser access, camera input, and business software through MCP servers, allowing spoken requests to trigger the same actions available from a terminal session.
A “second brain” adds context
A local CLI system called Base is tied into the workflow as a knowledge graph and memory layer. At the start of a session, on each prompt, and before tool calls, it matches the current task against stored rules, decisions, and prior lessons, then injects that context into the workflow. The model does not retain memory on its own; recall is handled by the graph and supporting memory files.
Main business tools are already connected
Daily-use integrations include Gmail, a workspace CLI for email handling, GoHighLevel for CRM and pipeline management, Skool for community operations, Read.ai, and Slack. Analytics and creative tools are also attached. The practical difference is that a business owner can issue a spoken instruction and receive a completed action or report without manually opening each platform or re-explaining the task.
Setup is relatively simple
Installation is described as a two-step process: install the Jarvis repository locally, then connect a voice model. Cloud voice options include Fish Audio and ElevenLabs, with Fish Audio noted as offering inexpensive API usage and a prebuilt Jarvis-style voice. An MCP connection to the voice provider can also be added so the assistant can manage more actions inside that service, including voice selection and cloning.
Permissions determine how much Jarvis can do
Out of the box, the system runs through headless Claude Code sessions and may only read limited information. Full terminal-like behavior requires broader admin access and the attachment of local rules, memory hooks, and second-brain files. The goal is to make the voice assistant function like the existing terminal workflow rather than as a separate lightweight chatbot.
Email triage is one of the strongest early use cases
In one demonstration, the assistant reviewed the previous 24 hours of email, identified priority conversations, and drafted responses in batches of five threads at a time. It surfaced 48 emails across two inboxes, summarized the most urgent matters, and then carried out a follow-up instruction to send a selected reply while ignoring the first four drafts. It also created a task in the memory system tied to the issue.
Pipeline management can be turned into a spoken briefing
Connected to GoHighLevel, the assistant analyzed a CRM stage and reported 47 real leads alongside 8 test contacts. It flagged 14 contacts as worth emailing and identified a smaller set as immediate call priorities, including leads such as Eton, Nico, Kate, Justin, and Rob. It then converted that shortlist into an Apple Notes entry using an Apple Bash tool.
Invoicing can be produced from existing templates
For invoicing, the assistant used a prebuilt template and read an existing email thread to infer the needed details. It generated an invoice for Fish Audio contact Evan, applied a branded CC Strategic Brand Kit, and saved the finished file to the Downloads folder. The result matched the output expected from the terminal-based workflow, showing that repetitive back-office tasks can be delegated with minimal back-and-forth.
The broader appeal is time savings for owners
The most common automation requests gathered from more than 300 business owners were email triage, pipeline management, invoicing, and application building. The first three are presented as practical, near-term wins because they involve structured, repeatable tasks in tools many companies already use. The larger pitch is that an AI operating system of this kind can reclaim at least five hours a week by taking over routine operational work.
CONCLUSION
Voice-controlled AI agents are moving beyond novelty when they are tied directly to business software, memory systems, and repeatable workflows. The main test now is not whether they can perform tasks, but whether their access, reliability, and context handling are strong enough for daily operational use.
Explain this