kAIs is a Kubernetes-native multi-agent platform. The CLI installs the operator into your cluster,
kais login projects your local tokens, and a single kais run streams a
thinking-loop response back over A2A — with N parallel workers reading the same Context if you ask
for them.
$ kais init # apply 7 CRDs to the current kubeconfig ✓ applied CRDs (cells, formations, missions, contexts, connectors, connectortypes, artifacts) $ kais login -n demo # scrape gh / aws / ANTHROPIC_API_KEY → Secret Detected credentials: ANTHROPIC_API_KEY ← llm GH_TOKEN ← github AWS_ACCESS_KEY_ID ← aws Applied Secret demo/kais-tokens (3 keys). $ kais apply -f reviewer.yaml cell.kais.io/reviewer created context.kais.io/project created $ kais run reviewer "review the last PR" --workers 4 --merge fork → task task_7f9a on demo/reviewer [thinking] reading PR diff… [tool gh_pr_view] {"number":42} [result] +183 −62 lines across 8 files PR looks safe to merge — see review on GitHub.
Everything you need to deploy, address and reason about agents on a cluster you already control.
Every Cell pod stands up a Google A2A server on :7778 and an MCP server on :7777 — Claude Code, Cursor and other agents talk to it without a SaaS in between.
--workers N --merge readonly|fork|append spawns N sibling tasks against the same Context snapshot and multiplexes them onto one SSE stream.
Working memory and KV state live in a snapshot-isolated Context store (SQLite by default, Postgres opt-in). CRDT log keeps concurrent writes consistent.
kais login reads gh auth token, AWS, gcloud and env vars and projects them into a namespace-scoped Secret. Tokens never leave your machine.
GitHub, GitLab, Slack, Postgres, Neo4j, S3, SSH — defined as ConnectorTypes, instantiated as Connector CRs, surfaced as MCP tools inside a Cell.
One controller binary reconciles Cells, Formations, Missions, Contexts, Connectors and Artifacts. No managed-SaaS dependencies, no helmfile.