Products Decoded · Open source

Holt

Everything you know, kept and connected.

An open-source, self-hosted personal agent OS. Clone it, pick your skills, choose your LLM, and it runs on your own machine with persistent memory you can actually see and navigate.

npm i -g @holt-os/holt

Early stage and open source. This is a project you can follow, run, and shape while it is being built.

What it is

An assistant that remembers, and shows you the connections.

Most assistants forget you the moment a session ends, or they keep a private profile you never get to read. Holt does the opposite. It keeps a private, navigable knowledge graph of what you know, on your own machine, under your own keys.

You choose the model, from Claude to a local one. You choose the skills. A local model does the work, and a cloud model steps in only to review the steps that actually matter. Nothing leaves your machine unless you decide it should.

It is open-core and MCP-native, so the whole thing is yours to extend.

Why Holt

Six things that make it different.

01
Memory you can see

Most assistants keep a black box profile of you that you cannot open. Holt keeps a navigable knowledge graph of what you know, so you can walk the connections yourself instead of trusting a hidden summary.

02
Any LLM, no lock-in

Claude, OpenAI, Gemini, or a local model. Swap the brain with one line of config. The same skills, memory, and hooks run underneath whichever one you pick.

03
Local executes, cloud reviews

A local model does the day to day work for free. A cloud model is called in only to review the risky, irreversible steps like writing files, sending messages, or spending money. Cheap and private by default.

04
An MCP plugin pantry

Skills, channels, providers, and embeddings are all plugins that speak MCP. Extend Holt in any language that can speak the protocol, without touching the core.

05
Standard skills

Skills follow the agentskills.io format, so anything written for that standard works here, and anything you write here travels elsewhere.

06
CLI first

It lives in your terminal. Run it, talk to it, script it. Channel adapters like Telegram plug in when you want them, but the command line is home.

Quickstart

Three commands to your own agent.

1
Install
npm i -g @holt-os/holt

One global install. Node is all you need to start.

2
Run
holt

A first run wizard sets up your brain, your memory, and where embeddings live.

3
Talk
holt chat

Chat from the terminal. Your memory grows as you go.

Prefer to read the code first? The whole thing is on GitHub, MIT licensed.

Where it stands

The build, in the open.

Holt is early. The core is coming together and the roadmap below is the full plan, built in phases that each stand on their own. Watch the repo to see it land.

Shipping
The core loop

The agent runtime, brain router, MCP plugin dispatcher, and the holt chat command. Talk to a model from your terminal.

Shipping
Persistent memory

Local vector store, local or cloud embeddings, and session hooks that remember across runs. The graph starts here.

Next
Any LLM and output toggle

Provider plugins for Claude, OpenAI, Gemini, and local models, plus an HTML or Markdown output switch.

Next
The skill catalog

Install, search, and publish skills in the agentskills.io format. Run skills written in any language.

Planned
The knowledge graph view

The headline. A surface where you can actually see and navigate your own memory.

Planned
The review gate

RAM aware local executor, cloud reviewer, and an approve, edit, or block step for anything risky.

Follow the build

If a private assistant with memory you own sounds right, star the repo and follow along.

A star is the easiest way to help early, and it is how you hear when each phase ships. Contributions are welcome the moment you want to write a skill or a plugin.

npm i -g @holt-os/holt