graph-agents-cli

Graphs in, agents out.

Build, evaluate and deploy LangGraph agents on your own Kubernetes, with one CLI and six skills for your coding agent.

imagePyPI imagePython imageCI imageLicense imageDocs imageSkills tuned with SkillOpt imageDemo video

Get started · Guides · Reference · Skills benchmark · Changelog · Issues

graph-agents-cli is a port of Google's agents-cli to LangGraph on self-hosted Kubernetes: it keeps the lifecycle and the scaffold engine, and replaces ADK on Google Cloud with LangGraph, Helm and any cluster (how the two compare).

Works with your coding agent: Claude Code · Codex · Gemini CLI · Cursor · Antigravity · and others. Or run each command yourself: the CLI works without a coding agent.

What's new

• 2026-09-29 · 0.3.1 is on PyPI. The first release published there: uv tool install graph-agents-cli (release notes).
• 2026-09-29 · 0.3.0: agents calling agents. Agents ask other agents over A2A for the user they serve, a person's approvals stay with that person, and system checks, wires and deploys several projects as one (guide). Also new: structured final answers in a JSON shape you declare, and skills tuned with Microsoft's SkillOpt (results). Measured while building it (0.3.0 notes):
  • wiring an agent to five other agents takes 5 peer add commands instead of 40 api commands and 418 hand-written lines;
  • replies that were not a bare JSON document (gpt-5-mini, 24-case triage task) went from 28 of 48 to 0 of 96 with structured answers.

What you get

Nothing is specific to one domain or one company: a project picks its auth policy, declares its APIs and configures the rest through environment variables and chart values.

Demo video

https://github.com/user-attachments/assets/ebc36951-e765-490e-8c45-d412f1cecba7

Get started

You need Python 3.12 or 3.13 and uv. setup uses Node.js (npx skills) when it is there; deploying also needs helm, kubectl, git and a docker that builds with BuildKit.

1. Install the CLI and the skills

uv tool install graph-agents-cli
graph-agents-cli setup      # optional: add the six skills to the coding agents it finds

setup also reinstalls the same release from its git tag (Installation & setup explains why).

Other ways to install: pipx, the release tag from GitHub, extras

pipx install graph-agents-cli
uv tool install git+https://github.com/ss7172/graph-agents-cli@v0.3.1   # the same release, from its git tag
uv tool install 'graph-agents-cli[a2a,langsmith]'                        # adds run --mode a2a and eval submit

Mirrors and disconnected installs: Installation & setup.

2. Run your first agent, no model key needed

graph-agents-cli create my-agent && cd my-agent
cp .env.example .env
export MODEL_PROVIDER=fake            # the keyless, deterministic test model
graph-agents-cli login --write-env    # checks the setup, generates API_KEY
graph-agents-cli install
graph-agents-cli run "What's the weather in San Francisco?"
graph-agents-cli eval run             # the exit code is the gate

For a real model, leave out the export line and login --write-env asks for the key. The provider is one setting: openai, anthropic, gemini, or openai-compatible for Ollama, vLLM, TGI and others. The Quickstart walks through each step.

3. Build with your coding agent, then ship it

Once setup has installed the skills, open your coding agent in an empty directory and ask, for example:

"Use graph-agents-cli to build an agent that answers questions about orders from our orders API (GET /orders) and can put an order on hold (PATCH /orders/{order_id}). Holding an order must wait for my approval. Use the fake model locally; when the evals pass, deploy it to my local kind cluster."

The skills lead each step and stop for your decision at every gate (coding-agent tutorial). Or do it by hand: add an API tool under a policy, pass the eval run gate, then deploy --env dev to a local kind cluster (manual tutorial).

Skills, measured

setup installs six skills, all measured on gac-bench: 104 realistic graph-agents-cli tasks with deterministic verifiers. SkillOpt proposed edits to the workflow, scaffold and observability skills, and every edit was reviewed by hand before it shipped.

0.3 skills against 0.2 skills on gac-bench (hard pass rate, same CLI build, only the skill text differs):

Codex ran the test split of all six skills plus val of workflow and observability. Change is the mean of per-task differences, so it need not equal the difference of the two scores.

Where the gain is, per skill (Claude Code, val and test tasks)

The other four skills scored 0.92 or higher before and moved by 0.00 to +0.06.

No test task went down on either harness. There is one real regression on Claude Code (1 of 2 reps of a new deploy task in the val split), and the test-split gains alone are not significant (p = 0.25 and 0.5). Full tables: final-v0.3.md.

Commands

All commands

graph-agents-cli <command> --help prints the same reference as the site's CLI page.

Documentation

Status

Version 0.3.1, alpha, on PyPI and as tags on GitHub (release notes). Interfaces may still change between minor versions; the changelog lists every breaking change with its migration steps.

0.3's acceptance tests covered a system of six agents built with peer add and system apply on a local cluster, security probes, a probe of 20 agents with 2 replicas each, an issuer that hangs, and an upgrade from 0.2.0 with data. Blocker and major issues are fixed before a release; the rest are parked.

• Known issues: parked issues, each with its impact and a workaround where one exists.
• Changelog: every release and its migration steps.
• Contributing: development setup, tests, templates, releases and the upstream-sync process.

Credits

graph-agents-cli stands on two projects:

• google-agents-cli, Copyright 2026 Google LLC, Apache-2.0. graph-agents-cli is a fork of it, started from 1.6.1, with the Google Cloud specific parts removed; files kept from it keep their Google LLC copyright headers. NOTICE lists the modifications.
• SkillOpt from Microsoft, which found the skill rules that 0.3 adopted after review.

License

Apache-2.0: see LICENSE.