Alpha preview: pin 3.0.0a3 when you install. A bare pip install agentbreeder still installs the unmaintained 2.x: known issues.
agentbreeder

Console (agentbreeder up)

agentbreeder up starts a loopback, single-user web console (formerly the command center) that runs the same CLI commands to build, release and inspect agents.

Local, single-user preview

The console runs on 127.0.0.1 for one person on one machine. Its screens were run on macOS with Docker Desktop, mostly through Playwright and the console's own API, from a wheel built from main. The published 3.0.0a3 has been run through sign-in only (scripted, no browser). Linux has not been run, and 3.0.0a2 called it the command center. Each screen below says what has not been run.

npx agentbreeder@3.0.0-alpha.3 up                              # needs Node 18+ and uv
uvx --from "agentbreeder[eval]==3.0.0a3" agentbreeder up       # needs uv

The cluster steps also need kind 0.33 or newer and kubectl on PATH. It prints a one-time sign-in link and opens it (5 minutes, single use; press Enter for a new one). --port, --no-browser and --environment are available. Stopping it leaves services and deployments running.

Every action runs a CLI child process, so the identity gate, policy, eval gate and audit log all still apply. There is no approve button: a second person's verified identity is what a single-user session can't provide.

Screens

AreaWhat it does
HomeThe guided path and Create starter; Build an agent from a form with a live preview of every file; onboard an agent.yaml alone; the services table with CPU, memory and IO; LLM cost; container events and failed jobs
AgentsEvery agent on this machine, and each agent's gate strip (Build, Candidate, Evals, Identity, Production) with Overview, Try it, Releases, Logs and Traces tabs
Try itChat with production. Turn details show tools, sources, latency, tokens, cost and the trace
RegistryBrowse, diff and see used-by; edit prompts, eval cases and knowledge documents as new versions; import and fork, with a confirm for code
MCP grantsRelay state, pins, drift and grants; grant, revoke and repin after a review
ServicesStart, stop and restart one service, its logs (with Follow) and usage
GovernanceThe audit log with paging and agent links, and saved eval reports
KubernetesConnect a local kind cluster and deploy the released image to it (Kubernetes on kind)

What has been run

  • Starter to reply: Create starter, add the key, review and start services, Build, Run release, Try it (first reply about 2 minutes in), and the Kubernetes step.
  • Build an agent: a Gemini agent with a prompt and evals, and one with a tool after the code confirm, each released 1/1 and answered in Try it. Create stays off for combinations not run end to end, such as MCP servers or two tools.
  • Northwind: onboarding, the grant review, release 61/61 and promote 61/61 through the console's API, and every screen the console had on 2026-11-13 opened in Chrome.
  • Editors: prompt and eval-case editing, Move agent to this version, Build and update agent.lock, the gate-failed view, knowledge documents, MCP grant, revoke and repin.
  • Inspect: Agents, Try it, Traces, LLM cost, the eval report viewer, Audit paging, the registry browser, import and fork, service restart and usage. Logs, Follow and container events ran through the console's API, not clicked.

Not run against a live stack: the tool code editor, tool detail and tool log screens (tested in CI), the Docker-down notice, and how Try it shows runtime/adk@1.3.0 and 1.4.0 model error codes, fallbacks and cached tokens (tested with a fake CLI). No screen has been clicked through on Linux.

Limits

  • Whoever controls local Docker has full control; the gate prevents mistakes and policy drift, not a local admin.
  • Try it calls an unauthenticated endpoint and costs real model calls.
  • Sessions live in memory: a console restart ends them.

Canonical reference: command-center.md.

On this page