Private AI, running on your phone

The personal AI system that answers only to you.

Local chat, vector memory, optional web search, voice, skills, and MCP tools in a phone-first assistant designed for privacy.

Download for AndroidView GitHub
New  MCP tool routing ->
Munin

CURRENT CHAT

Research local LLM memory

I run entirely on your phone. I can use local memory, search the web when you ask, and route agent work through tools you approve.
Find what I said about vector memory and draft a short architecture note.
Reasonedlocal only

Search memory for semantically close chats, attach the top results, then draft without sending chat history anywhere.

I found 4 relevant memories and drafted a note around SQLite vector recall, profile facts, and per-turn context injection.
Ask Munin anything
SearchThinkAgent

Local inference

Gemma on-device

Vector memory

SQLite recall

Optional web search

Explicit network

Voice mode

Hands-free chat

MCP tools

Approved routes

Open source

Inspectable app

Private by default

Local memories

Agent skills

Tool-aware work

A new class of personal assistant. Local-first at the core, capable enough to help with the real world.

FIG 0.1

Private by default

Messages, memories, and settings live on-device. Munin only reaches the network for actions you choose.

FIG 0.2

Built around memory

Semantic recall makes old conversations useful without turning them into a cloud profile.

FIG 0.3

Tool aware

Search, voice, device actions, SKILL.md files, and MCP servers share one visible agent surface.

Run a capable model without handing over your life.

Munin is designed around on-device inference with Gemma, local SQLite storage, and a privacy boundary you can actually understand.

1.0 Local engine ->

munin/runtime
local session
$munin status
>model: Gemma 4 E2B
>backend: LiteRT GPU
>context: 4096 tokens
>network: offline
>privacy: no telemetry
device fit
adaptive
CPU44%
GPU78%
Memory56%
1.1 Gemma runtime1.2 Offline replies1.3 Model manager

Search every past conversation by meaning, not keywords.

Munin embeds chats locally, retrieves the closest memories, and injects only the useful context back into the next reply.

2.0 Vector memory ->

conversation index
sqlite vec0
0.94Architecture note: vector memory should be local
0.89Preference: keep memory visible and wipeable
0.82Prototype: inject top matches into each turn
0.77Search notes: do not upload chat history
semantic space
top-k recall
query vector
01
embed
02
rank
03
inject
2.1 Conversation vectors2.2 Similarity recall2.3 Local SQLite

Turn a request into a small plan of action.

Workspaces break bigger goals into steps, use tools where needed, and leave files and logs behind for review.

4.0 Workspaces ->

workspace run
agent steps
Research project constraints01
Create a draft plan02
Ask for missing details03
Export review checklist04

Run surface

voicesearchintentjsmcpfiles

Each step is explicit, inspectable, and gated by the permissions you grant.

4.1 Device actions4.2 SKILL.md4.3 Run logs

Connect tools without breaking the privacy boundary.

Munin can route agent work through MCP servers and local skills with explicit tool surfaces, visible permissions, and inspectable output.

5.0 MCP ->

mcp routing
permissioned tools
Munin
MCP
Filesystem
MCP
Browser
MCP
Notes
MCP
Calendar
approval sheet
before execution

Allow MCP call?

Munin wants to ask the filesystem server to create a local markdown note.

server: filesystem
method: write_file
path: notes/vector-memory.md
Deny
Allow once
5.1 MCP servers5.2 Permission gates5.3 Custom skills

Built in the open, designed to be inspected.

Munin is MIT licensed app code with a clear privacy model. Fork it, audit it, change it, or run it exactly as shipped.

MIT
App license
Flutter
App stack
Gemma
Local model

Private by design. Available today.

Start with local chat and keep building toward memory, search, voice, and agent tools without creating an account.

Download for AndroidGitHub