Open source Self-hosted · AGPL-3.0

A self-hosted workspace for AI agents

You give an agent a task, and it works on it in a sandboxed container on your server, with a shell, a browser, and document tools. Projects keep their files and task history, so you can reopen a task later and continue it.

Install Documentation

Works with any model provider that supports the OpenAI Responses API.

A Meowbert task: the agent has charted three months of bike rides, and the activity panel shows the shell commands it ran.

Who it's for

Personal projects, coursework, and client work

Meowbert runs on one machine with Docker Compose, supports multiple users, and stores everything in your own database. One person can run it on a home server, and a small team can share it on a VPS.

Hobbyists

Run it on a home server or a small VPS and connect whichever model you use, including one you host locally. Long tasks keep running after you close the browser.

Students

Keep a project per course with its files and tasks together. Agents can look up sources, write Word documents and slides, and run code in a sandbox rather than on your laptop.

Small teams

Workspaces can have several members, shared projects, and per-user limits, managed from an admin panel. Results can be sent to Telegram, Discord, email, or push notifications.

One Compose stackThe app, Postgres, Redis, Meilisearch, and a sandbox image.

AGPL-3.0The full source is on GitHub.

No telemetryNothing is sent to the Meowbert project.

Examples

Example outputs

These came from a fresh install and three short requests. The files are shown as the agent produced them; the chart took one follow-up message.

Four slides from a generated PowerPoint deck: a quarterly IT maintenance plan with a risk table, a four-quarter cycle, and next steps. — Draft a 6-slide PowerPoint proposing a quarterly IT maintenance plan for my client. PowerPoint, about 2 minutes

The first page of a generated Word document titled Urban Tree Canopy and Summer Surface Temperatures, with a summary and key findings citing published studies. — Find credible sources on tree canopy and urban heat, and write a two-page literature summary. Word document citing 4 published studies, about 2 minutes

A bar and line chart of weekly riding distance and elevation, with the longest ride called out. — Make a CSV of my rides, then a Python script that charts a weekly summary. Python in the sandbox, about 1 minute plus one follow-up

Features

Main features

Projects keep their files and history

Each project has its own files, instructions, memory, and task list. Tasks don't expire, so you can reopen one weeks later and send a follow-up. Each project also has a Master, an agent you can ask to start tasks, check on them, and report back.

Continue any past task
Branch a conversation to try a different approach
Workspace memory that agents search and update

A project's Master answering a status question and confirming a weekly scheduled review.

Each task runs in its own container

Tasks run in short-lived Docker containers with a shell, git, Python, Node, a browser, and Office and PDF tools. The agent can install packages and run the code it writes, and it can only reach that container and the project's files.

Read-only system filesystem, no Linux capabilities, CPU and memory limits
Optional gVisor runtime
Live terminal access to any project

A research task with its activity panel open, showing the agent's reasoning steps and the tools it called to build a Word document.

Scheduled and long-running tasks

Tasks can run on a cron schedule, loop until you stop them, or run for a set number of minutes. You can also start tasks and get their results through Telegram, Discord, GitHub, or email.

Step and time budgets
Push and email notifications
Queued and running work resumes after a server restart

The task composer's parameters menu: workflows such as Long Horizon and Agent Swarm, and scheduled, infinite, and timed tasks.

Any model with the Responses API

Meowbert connects to models through the OpenAI Responses API, so it works with OpenAI, API gateways, and local model servers that support it. Admins can offer several models with different reasoning effort, and an Agent Swarm has several models discuss a task before answering.

Reasoning effort set per model
Users can bring their own API key if the admin allows it
Long Horizon and Deep Research workflows with a reviewer step

The model picker in the task composer, listing hosted models alongside one running locally.

Other features

Also included

Documents and slides

Agents create Word, PowerPoint, and PDF files and render them to check the layout.

Interactive Canvas

A small website stored in a project that agents build and keep editing.

Browser automation

Agents can open pages, click through them, and extract content.

Deep Research

The agent plans its research, reads sources, and a reviewer checks the report before it's delivered.

Sources

Google Drive, OneDrive, pCloud, Outlook, rclone remotes, and YouTube.

Custom skills

Add a folder to skills/ to give agents your own tools.

Desktop apps

macOS and Windows apps with a quick-launch composer and computer use.

Installable web app

Install it on your phone as a PWA with push notifications.

See the documentation for the full feature list.

Status

Current status

Meowbert started as a personal project in February 2026 and was developed privately for about 7 months before this first public release. It's in daily use, but some parts are less polished than others.

Guarantees

The app, database, search index, and sandboxes all run on your hardware, and there's no telemetry.
Tasks and files are stored in your own Postgres database and on your own disk, and nothing expires.
The code is licensed under AGPL-3.0.

Known limitations

Model providers must support the OpenAI Responses API. Servers that only offer Chat Completions don't work yet.
Production use is only tested on Linux. Docker Desktop on macOS and Windows works for trying it out.
Sandboxes use Docker isolation, optionally with gVisor, but no sandbox is guaranteed escape-proof. Run Meowbert on a machine you're comfortable letting agents use.

All known limitations

Install

Installation

Clone the repository, run the setup script, and start the stack. The first account you create becomes the admin, and a setup wizard asks for your model provider.

Server: Linux, x86-64, 2+ cores
Memory: 4 GB minimum, 8 GB recommended
Software: Docker with Compose v2
Model: Any OpenAI Responses API provider

$ git clone https://github.com/XInTheDark/meowbert-ai-agent.git
$ cd meowbert-ai-agent
$ ./scripts/setup.sh
$ docker compose up -d

# then open http://localhost:5173

The quickstart guide covers access from other machines, adding users, and backups.

Meowbert is free and open source

Bug reports, feature requests, and pull requests are welcome on GitHub.

View on GitHub Documentation