Runtime v4 · Android-native agent · User-gated actions

A pocket agent
with cat-like
curiosity.

Meow Agent is not a decorative chatbot. It listens, reasons, checks permissions, uses registered Android tools, then reports what actually changed.

01 · Think

Analyze, reflect, plan

02 · Ask

Permission before risk

03 · Prove

Verify tool results

Download APK Star on GitHub

Meow Agent

Companion runtime online

verified

You

“Summarize notifications, add the important one to notes, remind me tonight.”

Runtime plan

Read notification snapshot

Ask before creating note

Verify note and reminder IDs

Control gate

“I need permission to create a note and reminder. Continue?”

modules

phases

runtime

langs

What makes it Meow?

Cute face,
serious runtime.

The character is not just the mascot. Meow Agent is designed like a careful companion: curious enough to help, restrained enough to ask, and honest enough to verify before saying it is done.

Companion layer

Understands the job

A request becomes structured intent, selected tools, a plan, execution steps, and a final explanation. It is built to work through tasks, not only answer text.

Analyze Reflect Plan Execute

Control layer

Permission gates are part of the personality.

Meow Agent does not pretend it can do everything. Sensitive actions are surfaced, Android permissions stay opt-in, and revoked permissions disable dependent behavior.

Present

Show what capability is needed.

Explain

Clarify why the action needs access.

Respect

Denied permission is a real stop condition.

Device-aware

Works through Android tools

Notes, files, calendar, notifications, calls, SMS, reminders, search, APIs, mini apps, and workflows are exposed as registered modules.

Local-first

Keeps app state on-device

SQLite stores core app data locally, while provider keys live separately in encrypted Android storage.

Bring your model

OpenAI-compatible providers

Connect your own provider, model, persona, and agent profile instead of depending on a hosted Meow server.

Verification-first

Done means checked

Mutating tools use result data or state re-checks so the final answer is grounded in what the app can verify.

Real runtime capture

One request.
The whole journey.

Watch Meow Agent understand a request, build a task plan, work through registered tools, and expose progress instead of hiding it behind a typing indicator.

Ask naturally

The runtime resolves intent and context.

Follow the plan

Each step stays visible while tools run.

See grounded results

Completion follows tool evidence and verification.

01:06 · End-to-end Android capture H.264 · 720 × 720

Inside the app

Built to feel calm.
Built to do work.

One visual language from conversation to permission gates, agent setup, modules, and multi-step workflows.

Swipe or drag to explore

Meow Agent home screen with active modules — 01
Home base
Modules and chat, one tap away.

Meow Agent chat showing a completed multi-step task — 02
Execution ledger
See the plan and what completed.

Meow Agent module store — 03
Module store
Install only what the agent needs.

Workflow module permission controls — 04
Permission controls
Every sensitive capability stays explicit.

Workflow template gallery — 05
Workflow templates
Start useful automations quickly.

Multi-step workflow builder — 06
Workflow builder
Chain agents, context, and failure rules.

Your agent

Choose its model, profile, and workspace.

OpenAI-compatible provider configuration — 08
Bring your provider
Connect an OpenAI-compatible endpoint.

Runtime, translated

What happens
under the paws.

Every visible response has an internal counterpart. The runtime understands, plans, acts through gated tools, verifies the result, then explains only what it can support.

Inside the runtime

The machinery

On the surface

What you experience

Analyze Reflect

Understand the actual goal

Resolve intent, relevant tool groups, missing context, selectors, and possible impact before choosing an action.

Understanding

“Let me carefully understand the goal and available context first.”

Plan

Turn intent into bounded steps

Build a concrete goal tree with ordered work, dependencies, confirmation boundaries, and completion criteria.

Task plan 0 / 3

Read the relevant device state

Run the requested action

Re-check the final state

Execute Verify Review

Act through controlled tools

Registry-owned risk metadata selects permission gates. Mutations must pass a verification probe before completion.

Permission gate

“Allow Meow Agent to create this reminder?”

Verification

Reminder ID returned and state re-read.

Verbalize

Report only grounded results

Turn structured execution evidence into a concise response in the user’s language, without inventing success.

Verified result

“Done. The reminder was created for Friday at 7 PM, and I confirmed it is in your reminder list.”

Shared foundation

The layers every phase relies on

Flutter and Riverpod present state; the router, plugins, local database, and encrypted storage provide controlled capabilities underneath.

ToolRouter

Dispatch plus exact permission gates.

ModulePlugins

16 self-registering capability modules.

SQLite

Local app state in meow_core.db.

Secure Storage

Provider credentials encrypted on-device.

Bounded recovery Cancellation Stuck-loop detection One bounded JSON repair Verification probes

Extensible System

Runtime Modules

16 self-contained ModulePlugins covering device integrations, productivity, and custom tools.

Device Context

Core

App launching, device info, system context

Notifications

Read, summarize, surface important notifications

Notes

Create, search, manage notes locally

Files

Browse, organize, share files

Calendar

View and manage events, reminders

Workflows

Scheduled, event-driven automation

Agent & Provider

Multiple agents, personas, memories

System Config

Settings, wallpaper, apps

Memory

Persistent cross-conversation memory

SQLite Queries

Safe read-only system database access

Calls & SMS

Initiate calls, send messages

HTTP APIs

Connect to custom APIs, reusable endpoints

User Databases

Custom data structures

Mini Apps

WebView apps backed by custom data

Reminders

Time and location-based alerts

Search

Web search, information retrieval

Quick Start

Get Started

Running locally or ready to contribute? Here's how to get started.

Prerequisites

Make sure you have these installed on your machine:

# Required tools

• Flutter SDK (Dart ^3.11.0)

• Java 17

• Android SDK

# You'll also need

• OpenAI-compatible API key

Install Dependencies

flutter pub get

Run on Device

flutter run

First Setup

On first launch, create a provider and agent:

# 1. Create a provider

• Provider nickname

• OpenAI-compatible base URL

• API key

• At least one model

# 2. Create an agent

• Assign provider & model

• Set persona & profile

Install Modules

After setup, install only the modules you need and explicitly grant their permissions. Every permission remains opt-in and revocable.

A pocket agent with cat-like curiosity.

Cute face,serious runtime.