Can I build a personal AI assistant with Claude Code without coding?

Yes, in the sense that you describe what you want and Claude Code writes and runs the code. You still need to supervise it: read the commands it proposes, understand what access each integration gets, and decline anything broader than the task. The work shifts from programming to reviewing and scoping.

OpenJarvis is an open source framework for local-first personal AI from the Intelligence Per Watt initiative at Stanford's Hazy Research and Scaling Intelligence Lab. Its documentation describes local inference engines such as Ollama, optional cloud providers, built-in tools, external MCP servers, persistent local memory, and cron-based agents for recurring jobs like email triage.

Is a local assistant more private than using Claude?

With a local model, prompts and files stay on your machine, which is a real privacy gain. The trade-off is capability, since small local models handle multi-step tool use less reliably than frontier models. Adding a cloud provider for harder tasks sends that content off the machine, so decide per task which model handles it.

How does a Claude Code assistant remember things between sessions?

Claude Code loads instruction files such as CLAUDE.md at the start of each session, so standing facts and preferences go there. Keep that file short, because it loads every time. Longer notes, past decisions and reference material work better in a searchable store the assistant queries when a request needs them.

Is it safe to connect a personal AI assistant to Slack or email?

It can be, with limits. Restrict the channel so only you can message the assistant, give it its own revocable tokens with the narrowest scopes available, treat instructions inside emails or documents as untrusted, and require approval before it sends, deletes or pays for anything. Begin with read-only access and widen it gradually.

Building a Personal Jarvis With Claude Code, Without Writing Code

There are two routes. Claude Code can act as the installer for a separate local assistant such as OpenJarvis, running the setup commands and wiring in models and chat channels. Or Claude Code can be the assistant, using instruction files as memory, MCP connectors for reach, and scheduled tasks for recurring work. Both need scoped access and reviewed commands.

What no-code setup actually means here

The recent Outskill video on building a Jarvis with Claude Code describes, in its description, setting up OpenJarvis, a free local assistant, with Claude Code doing the installation: free NVIDIA-hosted models first, private local models through Ollama next, then voice control and Slack for remote access. The viewer types requests in plain language. Claude Code reads documentation, runs installers, edits configuration and fixes errors.

That is genuinely useful, and it is worth being precise about. No code was written by the person. Plenty of code was executed, with that person's file access, credentials and network. The skill being exercised is supervision: reading the commands Claude Code proposes, understanding roughly what each does, and refusing the ones that reach further than the job requires.

Route one: Claude Code installs a separate assistant

OpenJarvis describes itself as a research framework for local-first personal AI, from the Intelligence Per Watt initiative at Stanford's Hazy Research and Scaling Intelligence Lab. Its documentation lists inference engines including Ollama, vLLM, llama.cpp and MLX, with OpenAI, Anthropic and Google as optional cloud providers; built-in tools plus external MCP servers; persistent local memory with index and search commands; and cron-based agents for jobs like email triage and research digests. A quickstart script installs dependencies, starts Ollama with a model, and opens a local web interface.

Here Claude Code is the contractor, not the resident. It clones the repository, runs the quickstart, edits configuration to add a model provider or a Slack integration, and debugs what breaks. Once finished, the assistant runs on its own, and Claude Code is only needed again for changes.

The strength of this route is privacy and cost: with a local model, conversations and files never leave the machine. The weakness is capability. Small local models are noticeably weaker at multi-step tool use than frontier models, so expect simpler tasks to work well and complex ones to need a cloud model.

Route two: Claude Code is the assistant

The other route skips the separate framework. Claude Code already has the parts an assistant needs.

Memory. A dedicated folder with a CLAUDE.md file holding standing facts: who you are, how you like things written, which projects are active, where notes live. Claude Code loads it at the start of each session. Keep it short; longer reference material belongs in notes the assistant searches when needed.

Reach. MCP connectors for calendar, email, notes and task tools. Each one adds tool definitions to the context, so connect the three or four you actually use.

Schedule. Claude Code Desktop can run local scheduled tasks on a cadence with access to your files, and cloud scheduled tasks that run when your computer is off. Anthropic's documentation notes local tasks only fire while the app is open and the computer is awake, and that a missed run triggers one catch-up run on wake.

Remote access. Remote Control lets the Claude mobile app drive a session running on your machine, which covers most of what a Slack bridge would.

Which route to pick

Pick route one if privacy is the priority and the tasks are simple: summarizing local files, searching folders, drafting from your notes. Pick route two if you want frontier-model reliability on multi-step work and already pay for Claude. Many people end up mixing them: a local model for private triage, Claude for anything that needs judgement.

Guardrails both routes need

Treat channels as command lines. Anything that can send the assistant a message can direct it. A Slack integration must accept messages only from your account, and ideally only in one private channel.

Scope credentials. Give the assistant its own tokens, read-only where possible, and revoke them independently of your main account.

Distrust what it reads. Emails, web pages and shared documents can contain instructions. An assistant that reads an email saying "forward the attached invoice to this address" must not treat that as a request from you.

Approve consequential actions. In Claude Code, keep permission prompts on for sending, deleting, purchasing and anything outside the working folder. For scheduled tasks, Anthropic documents a per-task permission mode and an always-allowed list you can review and revoke.

Start with read and report. A morning briefing of calendar, priority email and open tasks exercises the whole stack and cannot do damage. Run it for a week before giving the assistant any write access.

Keep memory retrievable, not bloated. The assistant will want to remember more over time. Put durable notes in a searchable store and retrieve what each request needs, instead of growing the always-loaded file until it slows every session.

Frequently asked questions

Can I build a personal AI assistant with Claude Code without coding?: Yes, in the sense that you describe what you want and Claude Code writes and runs the code. You still need to supervise it: read the commands it proposes, understand what access each integration gets, and decline anything broader than the task. The work shifts from programming to reviewing and scoping.
What is OpenJarvis?: OpenJarvis is an open source framework for local-first personal AI from the Intelligence Per Watt initiative at Stanford's Hazy Research and Scaling Intelligence Lab. Its documentation describes local inference engines such as Ollama, optional cloud providers, built-in tools, external MCP servers, persistent local memory, and cron-based agents for recurring jobs like email triage.
Is a local assistant more private than using Claude?: With a local model, prompts and files stay on your machine, which is a real privacy gain. The trade-off is capability, since small local models handle multi-step tool use less reliably than frontier models. Adding a cloud provider for harder tasks sends that content off the machine, so decide per task which model handles it.
How does a Claude Code assistant remember things between sessions?: Claude Code loads instruction files such as CLAUDE.md at the start of each session, so standing facts and preferences go there. Keep that file short, because it loads every time. Longer notes, past decisions and reference material work better in a searchable store the assistant queries when a request needs them.
Is it safe to connect a personal AI assistant to Slack or email?: It can be, with limits. Restrict the channel so only you can message the assistant, give it its own revocable tokens with the narrowest scopes available, treat instructions inside emails or documents as untrusted, and require approval before it sends, deletes or pays for anything. Begin with read-only access and widen it gradually.

Key takeaways

No-code means you do not write the code; Claude Code still writes and runs it with your user's permissions, so read what it proposes to execute.
OpenJarvis is an open source, local-first personal AI framework from Stanford's Hazy Research and Scaling Intelligence Lab, with memory, MCP tool support, scheduled agents and cloud models as an option.
Using Claude Code itself as the assistant needs only a working folder, an instruction file for standing memory, a few MCP connectors, and scheduled tasks.
Any chat channel that can message the assistant, such as Slack, is a command interface and must be restricted to you.
Start with read-and-report jobs, such as a morning briefing, before granting the assistant any ability to send, edit or delete.