Graphs in, agents out.

Build, evaluate and deploy LangGraph agents on your own Kubernetes

One CLI, and six skills for your coding agent, take an agent from create to a hardened Helm release: a streaming chat API, shared-key or per-user auth, an outbound API policy, human approval of risky calls and an eval gate CI can enforce.

Get started View on GitHub

Now on PyPI: uv tool install graph-agents-cli, with skills tuned by SkillOpt.

Works with your coding agent Claude Code Codex Gemini CLI Cursor Antigravity and more

# install the CLI from PyPI, then the skills
uv tool install graph-agents-cli
graph-agents-cli setup

# create an agent and ask it a question: the fake model needs no key
graph-agents-cli create my-agent && cd my-agent
cp .env.example .env
export MODEL_PROVIDER=fake
graph-agents-cli login --write-env
graph-agents-cli install
graph-agents-cli run "What's the weather in San Francisco?"

Get started in three steps

Everything runs on your machine first; a model key and a cluster come later.

• Install the CLI and the skills

  One uv tool install graph-agents-cli from PyPI, then setup adds the skills to the coding agents it finds.

  Installation & setup

• Create and run an agent

  create a project, let login --write-env fill in .env, install, then run a prompt. Five minutes on the fake model, no key.

  Quickstart

• Evaluate and deploy it

  Add an API tool under a policy, pass the eval run gate, then deploy --env dev to a local kind cluster.

  Tutorial: manual workflow

What you get

A generic toolkit: nothing in the CLI or the generated project is specific to one domain or one company.

• Scaffold a real service

  create renders a LangGraph project with a streaming chat API, an A2A endpoint, an eval harness, a hardened Helm chart and GitHub Actions workflows.

  Develop your agent

• Run it locally

  run sends a prompt through a temporary local server; playground serves a dev chat page with reload. A deterministic fake model needs no key.

  Quickstart

• Evaluate with a gate

  eval run sends every dataset case to the agent, grades deterministic checks and LLM judges, and its exit code is the gate your CI enforces.

  Evaluation

• Deploy to any Kubernetes

  deploy --env dev|staging|prod with Helm, directly or through Argo CD pull requests. Local clusters (kind, k3d, minikube, Docker Desktop) need no registry push: any valid name, such as --registry localhost/dev, works.

  Deploy to Kubernetes

• Secure by default

  One auth policy on every surface (shared bearer, OIDC/JWT or your own), an outbound API allow-list and human approval of the calls you choose.

  Security & production

• Built for coding agents

  Six bundled skills teach your coding agent the same lifecycle, so you can ask it to "use graph-agents-cli to build ..." and review each step. They are tuned with SkillOpt on a 104-task benchmark: 0.84 to 0.97 on Claude Code against the 0.2 skills.

  Build with a coding agent · Skills benchmark

One lifecycle, from prototype to production

Each stage is a command, and each command's exit code tells a script or a coding agent what happened.

1. Create

   A project with its API, auth, policy, chart and CI.

   create scaffold enhance

2. Develop

   Write tools, declare the APIs they call, try it.

   run playground api lint

3. Evaluate

   Grade every case; the exit code is the gate.

   eval run eval compare

4. Deploy

   Build, apply the Secret, roll out with Helm or Argo CD.

   build secrets apply deploy

5. Operate

   Watch rollouts, decide approvals, upgrade the project.

   deploy --status approvals scaffold upgrade

Every change goes round again: develop, evaluate, deploy. The lifecycle explains each stage.

Where to next

• Get started

  Install, run your first agent in five minutes, then follow a tutorial.

  Get started

• Guides

  Authentication, the API policy, approvals, evaluation, deployment, secrets and more.

  Guides

• Reference

  Every command and flag, environment variables, the HTTP API, exit codes and skills.

  Reference