mng: build your team of AI engineering agents

installation:

curl -fsSL https://raw.githubusercontent.com/imbue-ai/mng/main/scripts/install.sh | bash

mng is very simple to use:

mng create           # launch claude locally (defaults: agent=claude, provider=local, project=current dir)
mng create @.modal          # launch claude on Modal (new host with auto-generated name)
mng create my-task          # launch claude with a name
mng create my-task codex    # launch codex instead of claude
mng create -- --model opus  # pass any arguments through to the underlying agent

# send an initial message so you don't have to wait around:
mng create --no-connect --message "Speed up one of my tests and make a PR on github"

# or, be super explicit about all of the arguments:
mng create my-task@.modal --type claude

# tons more arguments for anything you could want! Learn more via --help
mng create --help

# or see the other commands--list, destroy, message, connect, push, pull, clone, and more!
mng --help

mng is fast:

> time mng create local-hello  --message "Just say hello" --no-connect
Done.

real    0m1.472s
user    0m1.181s
sys     0m0.227s

> time mng list
NAME           STATE       HOST        PROVIDER    HOST STATE  PROJECT
local-hello    RUNNING     @local      local       RUNNING     mng

real    0m1.773s
user    0m0.955s
sys     0m0.166s

mng itself is free, and the cheapest way to run remote agents (they shut down when idle):

mng create @.modal --no-connect --message "just say 'hello'" --idle-timeout 60 -- --model sonnet
# costs $0.0387443 for inference (using sonnet)
# costs $0.0013188 for compute because it shuts down 60 seconds after the agent completes

mng takes security and privacy seriously:

# by default, cannot be accessed by anyone except your modal account (uses a local unique SSH key)
mng create example-task@.modal

# you (or your agent) can do whatever bad ideas you want in that container without fear
mng exec example-task "rm -rf /"

# you can block all outgoing internet access
mng create @.modal -b offline

# or restrict outgoing traffic to certain IPs
mng create @.modal -b cidr-allowlist=203.0.113.0/24

mng is powerful and composable:

# start multiple agents on the same host to save money and share data
mng create agent-1@shared-host.modal --new-host
mng create agent-2@shared-host

# run commands directly on an agent's host
mng exec agent-1 "git log --oneline -5"

# never lose any work: snapshot and fork the entire agent states
mng create doomed-agent@.modal
SNAPSHOT=$(mng snapshot create doomed-agent --format "{id}")
mng message doomed-agent "try running 'rm -rf /' and see what happens"
mng create new-agent --snapshot $SNAPSHOT

mng makes it easy to see what your agents are doing:

# programmatically send messages to your agents and see their chat histories
mng message agent-1 "Tell me a joke"
mng transcript agent-1

mng makes it easy to work with remote agents

mng connect my-agent       # directly connect to remote agents via SSH for debugging
mng pull my-agent          # pull changes from an agent to your local machine
mng push my-agent          # push your changes to an agent
mng pair my-agent          # or sync changes continuously!

mng is easy to learn:

> mng ask "How do I create a container on modal with custom packages installed by default?"

Simply run:
    mng create @.modal -b "--file path/to/Dockerfile"

Overview

mng makes it easy to create and use any AI agent (ex: Claude Code, Codex), whether you want to run locally or remotely.

mng is built on open-source tools and standards (SSH, git, tmux, docker, etc.), and is extensible via plugins to enable the latest AI coding workflows.

Installation

Quick install (installs system dependencies + mng automatically):

curl -fsSL https://raw.githubusercontent.com/imbue-ai/mng/main/scripts/install.sh | bash

Manual install (requires uv and system deps: git, tmux, jq, rsync, unison):

uv tool install mng

# or run without installing
uvx mng

Upgrade:

uv tool upgrade mng

For development:

git clone git@github.com:imbue-ai/mng.git && cd mng && uv sync --all-packages && uv tool install -e libs/mng

Shell Completion

mng supports tab completion for commands, options, and agent names in bash and zsh. Shell completion is configured automatically by the install script (scripts/install.sh).

To set up manually, generate the completion script and append it to your shell rc file:

Zsh (run once):

uv tool run --from mng python3 -m imbue.mng.cli.complete --script zsh >> ~/.zshrc

Bash (run once):

uv tool run --from mng python3 -m imbue.mng.cli.complete --script bash >> ~/.bashrc

Note: mng must be installed on your PATH for completion to work (not invoked via uv run).

Commands

# without installing:
uvx mng <command> [options]

# if installed:
mng <command> [options]

For managing agents:

• create: Create and run an agent in a host
• destroy: Stop an agent (and clean up any associated resources)
• connect: Attach to an agent

• list: List active agents
• stop: Stop an agent
• start: Start a stopped agent
• snapshot [experimental]: Create a snapshot of a host's state
• exec: Execute a shell command on an agent's host
• rename: Rename an agent
• clone: Create a copy of an existing agent
• migrate: Move an agent to a different host
• limit: Configure limits for agents and hosts

For moving data in and out:

• pull: Pull data from agent
• push: Push data to agent
• pair: Continually sync data with an agent
• message: Send a message to an agent
• transcript: View the message transcript for an agent
• provision: Re-run provisioning on an agent (useful for syncing config and auth)

For maintenance:

• cleanup: Clean up stopped agents and unused resources
• events: View agent and host event files
• gc: Garbage collect unused resources

For managing mng itself:

• ask: Chat with mng for help
• plugin [experimental]: Manage mng plugins
• config: View and edit mng configuration

How it works

You can interact with mng via the terminal (run mng --help to learn more).

mng uses robust open source tools like SSH, git, and tmux to run and manage your agents:

• agents are simply processes that run in tmux sessions, each with their own work_dir (working folder) and configuration (ex: secrets, environment variables, etc)
• agents run on hosts--either locally (by default), or special environments like Modal Sandboxes (--provider modal) or Docker containers (--provider docker). Use the agent@host address syntax to target an existing host.
• multiple agents can share a single host.
• hosts come from providers (ex: Modal, AWS, docker, etc)
• hosts help save money by automatically "pausing" when all of their agents are "idle". See idle detection for more details.
• hosts automatically "stop" when all of their agents are "stopped"
• mng is extensible via plugins--you can add new agent types, provider backends, CLI commands, and lifecycle hooks

Architecture

mng stores very little state (beyond configuration and local caches for performance), and instead relies on conventions:

• any process running in window 0 of a mng- prefixed tmux sessions is considered an agent
• agents store their status and logs in a standard location (default: $MNG_HOST_DIR/agents/<agent_id>/)
• all hosts are accessed via SSH--if you can SSH into it, it can be a host
• ...and more

See architecture.md for an in-depth overview of the mng architecture and design principles.

Security

Best practices:

1. Use providers with good isolation (like Docker or Modal) when working with agents, especially those that are untrusted.
2. Follow the "principle of least privilege": only expose the minimal set of API tokens and secrets for each agent, and restrict their access (eg to the network) as much as possible.
3. Avoid storing sensitive data in agents' filesystems (or encrypt it if necessary).

See our security model for more details.

Sub-projects

This is a monorepo that contains the code for mng here:

• libs/mng/

As well as the code for some plugins that we maintain, including:

• libs/mng_modal/
• libs/mng_claude/
• libs/mng_pair/
• libs/mng_opencode/

The repo also contains code for some dependencies and related projects, including:

• libs/concurrency_group: a simple Python library for managing synchronous concurrent primitives (threads and processes) in a way that makes it easy to ensure that they are cleaned up.
• libs/imbue_common: core libraries that are shared across all of our projects
• apps/minds: an experimental project around scheduling runs of autonomous agents

Contributing

Contributions are welcome!