Meru is a personal AI assistant that runs on a machine you control. It answers questions from your notes and files, remembers what matters about you, and uses tools such as web search, Gmail, your calendar and your Obsidian notes. The models run on your own machine, so no prompt goes to an AI company.
This page covers the main parts in a few minutes. Level 200 explains how each part works, and level 300 is the full design.
Amit Arora · September 2026 · the design, written before the code
The parts
Figure 1. Everything inside the dashed line runs on your machine: the models, your data and the assistant itself. Only the tools you allow reach the internet, and each call goes through dispatch first. The scheduler, new in v0.5, runs jobs on a schedule: a question, a skill or a script.
Built in Go. Meru is two small programs: meru, which you type questions into, and merud, which runs in the background and does the work. Go builds them into single files that run on macOS, Linux and Windows with nothing else to install. Compared with Python, that means a smaller footprint, easier installs, and room to run many Merus on one server.
A desktop app.meru-desktop puts the same conversation in a window: past chats down the left, the answer in the middle with the files it used, and a panel that shows which tools ran, what Meru remembered and who the tools contacted. A switch picks where Meru looks: your files, your mail and calendar, the web, or nothing at all. Drop a file on the window, or pick one from any folder, and Meru reads a copy of it with your question. It asks before a risky tool call with a card inside the answer, and once a chat has read mail or the web it asks before every such call, so words in an email can't send anything on their own. Its Settings screen turns each tool off or on, and sets whether it asks first, or asks before every call, and picks the folders Meru reads, and sets up the servers Meru runs for you, such as Obsidian, with a Fix button for one that stops working; a Setup screen walks a new user through it. Like meru it only talks to merud, which makes every change. It runs on macOS and on Linux, where it needs GTK 4 and WebKitGTK 6.0. On a Mac, a microphone button turns what you say into text that you check before you send it, and the sound never leaves your Mac.
A Mac installer. Each release has a disk image with "Install Meru.app", which sets a Mac up in eleven steps: Meru itself, Ollama and the models, the folders Meru reads, web search, your Obsidian notes, your Apple Notes, skills and commands, Gmail and Calendar if you want them, your name, and starting merud. Each step says what it does and what it downloads before it starts, and you can skip any step but your name.
A Linux installer. On Linux, tested on Ubuntu 24.04, one command installs meru and merud, then offers the desktop app, Ollama, setup and starting merud at boot, asking before each step. On a machine with an NVIDIA card, setup offers the largest answer model that fits the card's memory.
Local models. Ollama, a free program, runs open-weight models, models whose files anyone can download, on your machine. The default lite setup needs about 2 GB of downloads and runs on most computers; the full setup uses a bigger model for better answers and needs a stronger machine. merud keeps the models loaded, so answers start fast. It loads the big answer model in the background when it starts, and a question that comes early waits for that load. The desktop app lists the answer models we tried, with the commands that download them, and switches between them without a restart. You can also name sets of models in the config file, switch between them from the chat with /model, and compare how each one did with /usage by model. The answer model can also come from another model server, on your computer or on one in your own network. Meru never sends a question to the internet, and it asks you before any question goes off your computer. You can also let the small, quick model answer greetings and simple rewording, while every question of fact goes to the answer model, and keep a larger model to ask by hand.
Your files. You choose folders to index, such as your notes or documents, and Meru reads the notes, PDFs, code, Word and Excel files and CSV files in them. It adds up a spreadsheet's rows with its own code, so a total comes out exact. Meru indexes nothing until you list one, skips secret files such as .env, and keeps the index current as you edit. It searches your files by keyword and by meaning, and names the file behind each answer. A question for the web or for one of your tools, such as your calendar, skips that search. Email and calendar stay with their tools and never enter the index. If the embedding model or either task prefix changes, Meru rebuilds the vectors and keeps keyword search available while it works. An older database without a prefix fingerprint keeps its vectors when its model and vector size match; Meru records the current fingerprint.
Memory. Meru saves what it learns about you as small Markdown files, plain text you can open in any editor, sorted into folders such as me/, preferences/ and projects/. You can read, edit or delete any of them. It also writes a short summary of each conversation, so it can recall what you discussed last week.
Your chats. Each conversation is one plain text file under ~/.meru/sessions, with one line per question, answer and tool call, in a format called JSON Lines. You can read, back up or delete these files. They are the record: Meru's search database is rebuilt from them and your folders, and Meru keeps no copy anywhere else. Meru can read them too, so it can tell you what you talked about in any earlier chat.
Skills. A skill is a Markdown file of instructions for one kind of task, such as writing a report or explaining a topic. Meru ships with five (writing, explainer, web-research, file-research and mail-research), loads a skill only when a question needs it, and picks up any skill you add. One line in the config file turns a skill off. A skill can bring the tools it uses, such as web search, but only ones you have turned on.
Limits. A question gets 5 minutes and a capped length of answer. When Meru can't answer inside them, it says sorry instead of going quiet.
Honest about what it did. Meru tells the model what its own tools can change, and never to claim an action no tool took. When an answer still says "Done" or "I've moved it" and no tool ran, Meru adds a note under it: nothing changed on your computer. An answer that says "I called the datetime tool" when no such call ran gets a note too, and so does a money amount that nothing Meru read in that turn holds, or a web link that no search or page in the chat gave, which the model may have made up. The model reads the time of day from its prompt instead of guessing it. Asked which model it is, the model calls a tool that reads the answer from Meru's own settings, where it would otherwise guess from its training.
Web search. Meru searches the web through SearXNG, a search engine that runs on your own machine, in Docker. Meru can run it for you, or use one you run. It needs no account and no key. Your search words go out to the search engines SearXNG asks; your question and your files don't. Meru can also read a whole public web page, or download a file, when the model asks. A page that builds its text with JavaScript, such as a job posting, opens in a hidden Chrome that Meru downloads the first time and runs only for that page. It asks you first before it opens a web address that neither a search nor you gave it, so the model can't send your data out hidden in an address. One setting turns page reading off. When you ask for the web, paste a link, or name something your files don't mention, Meru searches before the model answers, so the model reads what the web says about the thing you named instead of guessing from its training. Later questions in the chat remember what the web said. The model can search on any question, and its instructions tell it to search before it answers when it isn't sure of a fact, such as what a song means.
Tools and other agents. Meru uses tools through MCP (Model Context Protocol), an open standard many services already support: Gmail and Google Calendar, your Obsidian notes. It can also hand a task to another agent over A2A (Agent2Agent). Tools start switched off; you turn on the ones you want, and Meru asks you before any tool call that sends, deletes or changes something. To let it run a program on your machine, such as git log or GitHub's gh, you write the whole command in the config file and the model fills in only the blanks, such as which folder; no shell runs it. The config file carries read-only GitHub commands, switched off, and gh keeps its own login, so Meru never holds your GitHub token. meru setup offers two starter tools (Google and Obsidian), and merud installs, starts and restarts the ones you pick, and web search too, each at a version it pins. When a tool you added by hand stops, Meru leaves it stopped and tries it once more when your next question needs tools. That question also asks each running tool what it offers, so a tool you restart with new features offers them on that question. meru mcp shows which tools are connected. A tool on the internet that wants you to sign in, such as GitHub's or Slack's, opens its sign-in page in your browser when you click Sign in. API keys stay in a file only you can read, and never pass through the model.
Observability. Meru records how long each step takes, how many words the model read and wrote, and which tools it called. The Usage page charts tasks, tokens and answer times per day and lists the slowest tasks and the tools that failed, with nothing else to install. Meru takes these measurements with OpenTelemetry, an open standard for them, and hands them only to a program on your own computer, such as Prometheus, and only if you run one. Your questions and Meru's answers stay out of them unless you turn that on.
What happens when you ask
Figure 2. The model never calls a tool itself. It tells Meru what it needs, and Meru makes the call, after asking you if the call would change something.
Most questions skip the tool steps. When the model needs live information or an action, Meru calls the tool, and it only stops to ask you when the call would change something. The answer appears word by word as the model writes it.
The harness underneath
merud is an agent harness: the part that runs a model in a loop, hands it tools, checks each tool call and keeps the record. What makes Meru your assistant, such as searching your files, remembering you and going to the web first, sits on top of that harness. meru, meru chat and the desktop app talk to it, and so can a script, with meru run --json.
What Meru won’t do
Send a question to the internet, or off your computer without asking. Only an answer model you name can run on another computer, one on your own network, and Meru asks "this time" or "this chat" first; an incognito chat never leaves. It refuses the cloud models Ollama offers, which would run on ollama.com without asking.
Send crash reports or its measurements off your machine, or check for updates. It measures itself, and that data stays here. The one thing it sends about itself is an anonymous count a day, which you choose: its version, your system and chip, the model set, a rough install age, rough counts of your questions and of the days you used it, how many weeks in a row it has reported, and the version it reported last, with no ID and no text.
Turn on a tool you haven’t allowed, or run a risky one without asking you.
Run a command line the model wrote. It runs only commands you declared.
Every tool call Meru makes is written to a log you can read with meru log. An incognito chat keeps no record of what you asked; the log still names each tool it called.
Explore further
The roadmap lists the milestones in shipping order.