Local quickstart

Run your first agent review

Start with a $49 order. Let two agents try different prices in separate branches, review the changes, then verify a merge and undo. The example uses real MongoDB writes and needs no model API key.

You need Docker, npm or Homebrew, and Python 3.10+. The commands below use a macOS or Linux shell and a disposable local database.

01Install the CLI

Already installed Argon? Check the version below and continue to MongoDB.

Install the CLI with npm

npm install -g argonctl@2.1.3 --allow-scripts=argonctl

Installs the argon CLI with npm 11.16+ and a supported Node.js version. Allows only argonctl's binary installer.

npm 12 blocks package install scripts by default. The npm command above allows only argonctl to download its release binary. Use npm 11.16+ with a supported Node.js version, or choose Homebrew or the source build below.

Check the installed CLI

argon --version

Expected: version 2.1.3. The package is named argonctl; the command you run is argon.

Build from source instead

For contributors. Requires Git and Go 1.26.9 or a later patched release (Go 1.27 users need 1.27.2+). This puts the built CLI on PATH for the commands below.

Optional source build

git clone --branch v2.1.3 https://github.com/argon-lab/argon.git
cd argon
(cd cli && go build -o ../bin/argon .)
export PATH="$PWD/bin:$PATH"

02Start MongoDB

In terminal A, start a local MongoDB 7 replica set. Docker must be running; port 27017 must be available.

Terminal A · Start MongoDB

docker run -d --name argon-mongo \
  -p 127.0.0.1:27017:27017 \
  mongo:7 --replSet rs0

Initialize the replica set once the container is running.

Terminal A · Initialize the replica set

docker exec argon-mongo mongosh --quiet --eval \
  'rs.initiate({_id:"rs0",members:[{_id:0,host:"localhost:27017"}]})'

Expected: the response includes ok: 1. If MongoDB is still starting and refuses the connection, wait a few seconds and retry.

Already have this local container?

Use docker start argon-mongo to restart it. An already initialized replica set does not need rs.initiate again. These commands are for the tutorial database; use your own connection settings for an existing deployment.

03Open the console

Still in terminal A, point Argon at the replica set and check that it can capture document changes.

Terminal A · Check MongoDB readiness

export MONGODB_URI='mongodb://localhost:27017/?replicaSet=rs0'
argon doctor

Expected: all checks show PASS. If the replica set has no primary yet, wait for its election and run argon doctor again.

Terminal A · Keep the console running

argon console --no-browser

Open your local console on port 1818. Keep this terminal running while agents work: the process manages capture and sandbox expiry.

04Run the example

Open terminal B. Download the matching example source, then install the Python SDK from PyPI in a virtual environment. Git is required to download the example.

Terminal B · Get the example

git clone --branch v0.2.0 https://github.com/argon-lab/argon-agents.git
cd argon-agents

Terminal B · Install the Python SDK

python3 -m venv .venv
. .venv/bin/activate
python3 -m pip install "argon-agents==0.2.0"

Terminal B · Run the two-agent review

ARGON_API_URL=http://127.0.0.1:1818 \
  python examples/two_agent_review.py

A successful run prints:

reviewed_price: 44
conflicts: 1
restored_price: 49

The assertions verify identical pinned input, an accepted $44 proposal, a competing $1 conflict, and the restored $49 order. Find the printed project in the local console to inspect its history.

If the API connection is refused, check that terminal A is still running. An authenticated server also needs ARGON_API_TOKEN set in terminal B.

Read the complete example

Connect your own agent

Your local server is ready. Choose the interface your agent already uses.

Claude / Cursor Python / LangGraph REST API

History needs healthy capture and retained data. CLI-only workflows also need a running watch process and scheduled cleanup. Read the operating limits or the complete setup guide before connecting a real workload.