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.
/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)" brew install holt-os/tap/holt Homebrew pulls Node in for you.
/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)" brew install holt-os/tap/holt Homebrew on Linux pulls Node in for you.
winget install OpenJS.NodeJS.LTS npm i -g @holt-os/holt Homebrew is not available on Windows.
First time in a terminal? Follow the step-by-step setup guide → Made for non-coders, no commands to memorise.
Open source and self-hosted. A real command surface you can use today, still growing in the open.
Running a company? Holt Teams brings this to your whole team, on your own hardware.
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 your knowledge as files you own: a navigable graph and a cross-linked wiki, on your own machine, under your own keys.
You choose the model, from Claude to a local one. You choose the skills. Memory stays per folder and local, and a local model powers recall, so it works offline and nothing leaves unless you decide it should.
It is MIT licensed and speaks MCP, so your other tools can read its memory and the whole thing is yours to extend.
What makes it different.
Most assistants keep a hidden profile of you that you never get to read. Holt keeps yours as files you own: a knowledge graph you can walk and a cross-linked wiki you can read and edit. Nothing about you stays a black box.
Install one hook and Holt feeds the right past notes into every Claude Code prompt, then saves what mattered when the session ends. No command to run. Your memory grows while you work.
Claude, OpenAI, Gemini, Kimi, Grok, or a local model. Switch the brain in the middle of a conversation and the thread carries over. The same memory, skills, and hooks run under whichever one you pick. Gemini connects with an API key now that Google has retired its personal sign-in, and Holt walks you through it.
It runs in the folder you launch it from, on your machine, under your own keys. Memory is per folder and local, and a local model powers recall, so it works offline and nothing leaves unless you send it.
Thirteen skills ship in the box, and more live in a community registry you can search and add. Skills use the agentskills.io format so they travel, and Holt speaks MCP so Claude Code, Cursor, or Codex can read this folder's memory directly.
Run a task once, on a schedule, or as a named routine. Reach it from your phone over Telegram. Have it draft in your own writing voice. It works while you are away from the keyboard.
You never see a raw error. If something breaks, Holt logs the detail, tells you in one sentence, and holt doctor --fix repairs the usual causes on its own: it starts the memory engine, fetches what is missing, and mends a damaged config without losing a thing.
holt tour walks a first-timer from install to the moment it remembers something about them. One word updates it, one word resets a folder, and the help reads like a person wrote it. Built for people who have never opened a terminal.
An empty assistant gives generic answers, which is why most people install one and drift away. The onboard skill is a fifteen minute interview that fills its memory with your work, then does one real task with it. After that, audit shows what is still missing and level-up turns one manual chore a week into something Holt does for you.
One tool, a workspace of commands.
Everything runs from your terminal, in the folder you launch it from. Here is what you can reach for.
holt Start your assistant. It sets the folder up the first time, then launches your brain branded as Holt, with your memory wired in. A CLI brain opens its own full interface, a direct API brain runs inside Holt. One command either way.
holt init Step through setup on its own: trust the folder, choose and install a brain, sign in, and turn on local memory.
holt memory Per folder memory with semantic recall. holt memory add teaches it a fact on the spot, and holt memory global shares high value facts across your folders when you opt in.
holt wiki A cross-linked wiki an LLM keeps for you, derived from your memory. Turn on auto-sync and it stays current on its own, and it is always safe to rebuild.
holt graph See your memory as an interactive knowledge graph. Point it at your code and docs too, and it clusters them into communities with a god-node report.
holt hook Ambient memory for Claude Code: inject relevant notes into every prompt and capture facts when a session ends.
holt run Run one task non-interactively: recall, let the brain do it, then remember the result.
holt schedule Fire a task on a timer, using a real OS timer: launchd on Mac, cron on Linux, Task Scheduler on Windows.
holt routine A named job that bundles a task or skill with a schedule and where the output goes.
holt skill Thirteen skills ship built in, including onboard, audit, and level-up. Search and add more from the community registry, or create and publish your own. Portable agentskills.io format.
holt mcp Serve this folder's memory to Claude Code, Cursor, or Codex over MCP.
holt voice Teach Holt how you write, from a short interview or your own samples.
holt write Draft content in your voice, with checks that keep it from reading like a machine.
holt telegram Chat with Holt from your phone. holt notify pushes a message to you.
holt doctor Look at your machine and recommend the best way to run Holt on it. Add --fix and it repairs the common problems for you: starts Ollama, pulls the missing model, mends a damaged config.
holt tour A two minute guided first run: ask about your files, teach Holt one fact, watch it remember. Made for people new to the terminal.
holt update Get the latest Holt in one word. Detects npm or Homebrew installs on its own.
holt reset Start a folder over safely. Asks for a typed confirmation, and your own files are never touched.
Full command and configuration reference lives in the docs.
Four steps to your own agent.
/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)" brew install holt-os/tap/holt Homebrew pulls Node in for you.
/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)" brew install holt-os/tap/holt Homebrew on Linux pulls Node in for you.
winget install OpenJS.NodeJS.LTS npm i -g @holt-os/holt Homebrew is not available on Windows.
New to the terminal? Follow the step-by-step setup guide, made for non-coders.
holt init Trust the folder, choose a brain, and turn on local memory.
holt Launches your brain branded as Holt, with memory wired in. The wiki, hooks, and the rest build on this.
/skill onboard A fifteen minute interview that fills its memory with your work, then does one real task with it. Everything else gets better once it knows you.
Prefer to read the code first? The whole thing is on GitHub, MIT licensed.
The build, in the open.
Most of the plan is already shipping. What is left is the registry and richer views. Each piece stands on its own, and the repo shows it land in real time.
Bare holt launches the real interactive brain (Claude Code, Codex, or Gemini) branded as Holt, with setup and memory wired in. Any brain as a CLI or a direct API connection, and brain switching that keeps context.
Per folder memory, semantic recall on a local model, distilled facts, ambient hooks for Claude Code, and opt-in memory shared across folders.
Your memory as an interactive graph that now weaves in your code and docs with community structure, plus a cross-linked wiki an LLM maintains and can keep current on its own.
Run, schedule, routines, a skill registry to search and share, an MCP server, Telegram, a machine advisor, and voice aware writing.
No raw errors, ever: crashes log the detail and holt doctor --fix repairs the usual causes. A guided tour, one-word update and reset, and two new direct brains: Kimi and Grok, beside Claude, OpenAI, and Gemini.
An onboarding interview that fills memory with your work in about fifteen minutes, a weekly audit of what is still missing, and a level-up session that turns one manual chore into something Holt does. Scheduled jobs now install real OS timers on Windows too, not just Mac and Linux.
Voice and mobile ways to reach Holt, beyond the terminal.
What a full skill suite looks like: holt-jobsearch.
An open-source job-search agent that runs the whole loop on top of Holt. It builds your profile, scores a role on fit, tailors your resume, writes the cover letter, preps you for interviews, reads your inbox for replies, drafts the follow-ups, and keeps your pipeline board current. The core needs no API keys, and it drafts everything while you keep the submit button.
If a private assistant with memory you own sounds right, star the repo and follow along.
Install it and run it today, then star the repo to hear what lands next. Contributions are welcome the moment you want to write a skill or a plugin.
brew install holt-os/tap/holt npm i -g @holt-os/holt