Meru · personal AI assistant · written in Go · pre-alpha

An AI assistant on
your own machine

Meru runs open-weight models on a machine you control. You need no API key and pay for no tokens, so a question costs electricity, and the model on your disk today still runs a year from now. No prompt leaves the machine unless you allow it.

मेरु

In some Eastern traditions, Meru is the cosmic mountain that the sun, moon and stars turn around. This assistant takes the name because it works the same way: it stays in one place, on your machine, and your notes, tools and daily routine turn around it.

QR code to github.com/aarora79/meru
github.com/aarora79/meru
Architecture in three levels:
aarora79.github.io/meru/

01Everything runs inside the dashed linethe parts of Meru

YOUR MACHINE you meru chat, or one question starts in ~50 ms, streams the answer local socket to merud merud runs in the background scheduler jobs from config.toml agent loop routes, searches, answers dispatch checks each tool call, asks before risky calls, logs every call metrics (optional) Prometheus, local only local models · Ollama lite or full fast main embed your data plain files your folders notes, docs, PDFs memory files Markdown, by kind transcripts one per chat search index SQLite, rebuildable skills + yours writing · explainer · research other agents · A2A off until allowed tools · MCP + more Google · web search · Obsidian only tools you allow internet services search engine, Google Inside the line: the models, your data and the assistant. Black arrows stay on the machine. Coral arrows leave merud through dispatch, and only the tools you allow cross the line.

02Pick the profile your machine can holdlite is the default

TierJoblite (default)full
fastrouting, trivial answersMiniCPM5-2B, 1.6 GBsame
mainreasoning, choosing toolssame model as fastqwen3.6:35b-a3b-mxfp8, ~47 GB loaded
embedvectors for searchembeddinggemma-2:440m, text and imagessame
needshardware16 GB of RAM; a CPU is enougha 64 GB Mac, or a GPU with ~47 GB

Platforms: macOS on Apple silicon (tested), Linux, cloud virtual machines included (tested on Ubuntu 24.04 with an NVIDIA card), and Windows 10+ (untested at first).

03The rules the code keepsfrom the privacy boundary

■No prompt to the internet. A model you name on your own network gets a question only after your yes.
■One anonymous count a day. It holds no ID or text, and you can turn it off. No crash reports, no update checks.
■Deny by default. A new MCP server or A2A agent gives the model nothing until you allow tools by name.
■Loopback unless you say so. A remote server or agent needs remote = true in its entry.
■Every external action logged. Review them all with meru log.
!No sandbox. MCP servers run with your permissions, so add only ones you trust.
Pre-alpha, v0.7.0. The design came first, and the code follows it. On a Mac or on Linux, Meru answers from the folders you index and cites the files and pages it used.

04One turn, five steps, in a loop we wroteno agent framework · traced locally

1 · Route
fast picks a route from one token.
2 · Context
Skills, memories, hits, history, each on a budget.
3 · Main
main streams the answer to you, or asks for a tool.
4 · Dispatch
Checks, asks if needed, runs and logs each call.
5 · Repeat
Until the answer needs no tool, or 8 rounds.

merud traces each step with OpenTelemetry and hands the data only to a program on this machine, such as a Prometheus you run, with no prompt text unless you opt in. The Usage page charts it with nothing to install.

05Two searches, merged by rankBM25 + vec1 distance + reciprocal-rank fusion

Search covers your files, memories and past chats. Keyword search (BM25) finds exact strings such as code names; meaning search finds paraphrase. Go merges the two lists.

RRF SCORE = Σ 1/(60 + RANK) · RANKS INVENTED, SCORES COMPUTED 0.01 0.02 0.03 chunk A vec #1 · kw #1 0.0328 first in both chunk B vec #3 · kw #2 0.0320 never first chunk C vec #20 · kw #15 0.0258 lower in both chunk D kw #1 0.0164 exact code name, keyword only chunk E vec #1 0.0164 paraphrase, vector only

06Every tool call passes one functiondispatch · deny by default

allow?If not: denied.
→
confirm?If so: ask you.
→
callMCP, A2A or built-in.
→
tool_callsLog every outcome.
→
spanTime and result.
approve onceThis call runs. The next one asks again.
this sessionNo more asking about that tool until the chat ends.
denyThe call doesn’t run, and the model hears that you said no.

configure and schedule_job ask every time. A script has no one to ask, so dispatch denies; a job runs what its allow list names.

[[mcp.servers]]
name    = "obsidian"
command = "uvx"          # stdio child
allow   = ["search", "read", "append"] # rest: off
confirm = ["append"]       # ask me first

07Memory is a folder of Markdown files~/.meru/memory/ · one fact per file

me/           who you are
preferences/  how you like things done
projects/     work, goals, deadlines
people/       people you mention
reference/    where things live
other/        anything else
■Each file carries its date and source chat. Delete the file and Meru forgets it.
■The model saves one with the built-in remember tool, through dispatch.
■Each chat ends with a short summary, so “what did we decide last week?” finds it.
■Files hold the truth. meru.db, one SQLite file with keywords and vectors, rebuilds from them.

08Setup takes one commandmeru setup · skills · jobs

meru setup · first run

1Checks for Ollama and prints the install command if needed.
2Downloads the answer model that fits your Mac’s memory or your GPU.
3Writes config.toml and asks which folders to index.
4Offers three starter MCP servers (Google, web search, Obsidian). It adds each for you, or shows you what to paste.
5Answers a test question.

Skills and jobs

■writing, explainer and two research skills ship built in.
■Skills are Markdown in ~/.meru/skills/. Your edits win.
[[jobs]]
name     = "morning-brief"
schedule = "0 7 * * 1-5"  # weekdays 7:00
prompt   = "Brief me on today..."

09Shipping orderall shipped · each useful by itself

v0.1
It talks
First token under 1 s.
v0.2
It knows your files
Cites your notes.
v0.3
It can do things
Calls MCP servers.
v0.4
It knows you
Recalls last week.
v0.5
It acts unprompted
A brief, unasked.
v0.6
It runs on Linux
One-command install.
Pre-alpha, v0.7.0. On a 64 GB Mac the full set passed 89% of 50 test tasks, with a first word in 4.1 s. Panel 05’s ranks are made up.
QR code to github.com/aarora79/meru
Read the design, follow the build

Code: github.com/aarora79/meru — README.md, ARCHITECTURE.md, ROADMAP.md
Architecture in three levels: aarora79.github.io/meru/
Amit Arora · Apache-2.0 · Ollama on loopback · Go binaries you can copy to another machine and run.