MIT licensev0.1 · EarlySelf-hosted

Open-source tools / AxonCrew

NanoCrew.
A crew of your own.

A platform for running a crew of AI agents, each with its own identity, memory, tools and schedule. Agents work in Docker sandboxes, hand tasks to each other, and ask for approval before sensitive actions. You manage the crew from a web dashboard.

View on GitHub Read the setup

An early release: expect rough edges and breaking changes. Bring your own infrastructure and model-provider keys.

NanoCrew command center showing agent workspaces, schedules and an approvals inbox — Example dashboard from the NanoCrew README. The values shown are an example, not customer results.

01 / What it does

Separate roles.
Shared work.

Configure an agent around a job, give it the tools it needs, and keep the work visible as it runs.

A workspace for each role

Create an agent from a template or from scratch. Give it an identity, long-term memory, tools and a job.

A crew that can hand work over

Agents can message each other, hand off tasks and borrow tools. The dashboard includes a crew map.

Docker sandboxes

Agent cycles run in Docker containers with shell, files, a browser and a Python REPL inside the sandbox.

Human approvals

Sensitive actions, such as sending email, pause in an approvals inbox. Per-agent rules let you configure what requires review.

Schedules and triggers

Start work through dashboard chat, cron schedules, webhooks or the optional WhatsApp channel. Your deployment must be running.

Your choice of model

A LiteLLM gateway connects your providers with fallback models. Configure a daily budget; spend is checked after each model call.

Tools, skills and MCP

Built-in tools, on-demand Markdown skills and external MCP servers let you configure the tools available to an agent.

Work you can inspect

Review per-cycle logs, costs and run history, and see a live view of the agent's browser from the dashboard.

Roles you can configure

Examples, rather than promises of ready-made agents or guaranteed results.

Development

Provide a repository workspace and the tools for a coding task. Review the changes it produces.

Writing

Provide a brief, source material and the tools for drafting. Review the output before publishing.

Research

Provide a question and the web tools it needs. Review the findings and sources it returns.

02 / Under the hood

Your infrastructure.
A visible runtime.

The React dashboard talks to a TypeScript orchestrator. The orchestrator manages the agent containers, approvals, schedules and model relay.

Interface

Dashboard

React + Vite
Agents, chat, logs, files and approvals.

Control

Orchestrator

TypeScript + Express + SQLite
Registry, scheduling, budgets and cycle runner.

Execution / Docker socket

Agent containers

The Python nanoagent runtime runs the model/tool loop with an agent workspace and scoped credentials.

Models / relay

LiteLLM gateway

Model calls use the gateway and configured providers. Agents receive short-lived relay tokens instead of platform provider keys.

The orchestrator needs the Docker socket, which grants powerful host access. Review the deployment guide and security policy before exposing an installation.

03 / Run it yourself

From checkout
to your first agent.

Docker with Compose v2, Node.js 20+, Bash, at least one model-provider key and at least 20 GB of free disk for the initial image build.

Clone the repository

git clone https://github.com/shivasaini04/nanocrew.git
cd nanocrew
cp .env.example .env

02
Configure your local environment
In .env, set a provider key and a matching MODEL_PRIORITY. Set HOST_PROJECT_ROOT to the absolute checkout path, and choose ADMIN_EMAIL and a strong ADMIN_PASSWORD.
Generate separate values for SESSION_SECRET, SESSION_ENCRYPTION_KEY, nanocrew_SECRET_KEY and LITELLM_PROXY_KEY using the documented setup instructions.
03
Build and start
```
bash container/build.sh
npm install && npm run build
docker compose up --build
```
The first sandbox-image build takes time. When the dashboard API is listening, open http://localhost:3002 and sign in with your configured admin account. Compose binds the ports to localhost.
04
Create an agent and review its work
Open Agents, choose a template or start blank, select a model and send a message. Watch the cycle in Logs and review sensitive actions in the approvals inbox. Add a schedule or webhook when you want the deployment to trigger work automatically.

04 / NanoCrew & ChatGPT

Choose what
you want to operate.

NanoCrew gives you an MIT-licensed agent platform to run and configure yourself. Compare that operating responsibility with the assistant or hosted service you are considering.

Infrastructure

You run the orchestrator, Docker sandboxes, storage and updates.

Context and tools

You configure agent identities, memory, providers, tools and integrations.

Triggers and review

You set schedules, webhooks, budgets and the approval rules for sensitive actions.

For ChatGPT's current workflows and account-specific availability, read OpenAI's official guide. NanoCrew's repository establishes its own capabilities; it does not establish another service's current feature set.

05 / Keep exploring

The code and
the details are open.

Getting started

Install, create an agent and troubleshoot setup.

Architecture

Components, cycle lifecycle and data boundaries.

Configuration

Providers, model choices, limits and access.

Deployment

Host access, TLS, backups and updates.

Writing skills and tools

Add a playbook, tool or MCP server.

Contributing

Work on the project and report reproducible issues.

Watch the project launch film

Project archive

The NanoCrew launch film

Launch film · use the README for current release scope · Press play to watch. Sound and playback are in your control.