Codemowers Agent

Always thought that having an agent run commands on your personal laptop was nuts? Give it a workspace in Kubernetes.

codemowers/agent runs Claude Code or Codex in Kubernetes worker Pods. Each Task gets a persistent workspace for its Git checkout, build artifacts and CLI session. Tasks keep working while your laptop is closed: forge events, mail and alerts wake them, and you or a colleague can pick up the conversation from any browser.

Workers are confined by the cluster, hold only the credentials their Agent names, accept only authenticated events and push changes for human review. The security model covers who may change what, current limits and an example of separating Agents by trust level.

To try it without your own cluster, get a sandbox at trial.codemowers.cloud.

Features

• Claude or Codex. Each Agent picks a CLI and provider credential. A Task can switch to another model from that provider and keep its workspace.
• Persistent Tasks. Follow progress and queue prompts in the browser. The worker stays up for the Agent's idle timeout after each reply. Suspend stops it and keeps the workspace, session and queued messages; Stop interrupts the current turn only.
• GitHub and Gitea. Signed webhooks route issue and pull-request events to the Task that claims the issue URL. A TaskPolicy creates a Task when an issue is assigned to the Agent's forge account. Closing the issue completes the Task and stops its worker.
• Other events. Alertmanager, Slack and other senders post to authenticated receivers, and CEL turns each payload into a resource URI that selects the Task. Unmatched events go to a fallback Task.
• One wake per activity. The first event on a key wakes its Task. For the next three minutes by default, the same activity arriving another way, such as GitHub's mail after its webhook, is only recorded, and further events get one trailing wake. Set coalesceSeconds on an Agent, TaskPolicy or Task, up to a day for a busy issue reviewed daily.
• Mail. A Mailbox Pod holds one account's IMAP and SMTP credentials. Mail can wake Tasks, and workers send and read mail through its MCP server.
• Remote devices. Machines running device-agent connect over WebSocket; workers run commands and edit files on them through the Devices MCP server.
• MCP status. The Task header shows each MCP server's status, instructions and tools, and can disable one for that Task.

Compared with OpenClaw, Hermes and Google AX

OpenClaw is a personal assistant across messaging channels and devices, and Hermes focuses on memory, learned skills and personal automation. codemowers/agent is instead built for a fleet of coding agents on one cluster. Agents, Tasks, receivers and policies are custom resources, so the whole fleet can be declared in Git. Webhooks, mail and alerts wake the Task that owns an issue, thread or alert, and work goes from that issue to tested code, a pull request and a deployment, with humans merging. An idle Task holds no Pod, only a resource, a PVC and a queue, so one cluster can carry hundreds of Agents and thousands of Tasks fed by one forge webhook per organization.

Google AX runs on Agent Substrate, with PostgreSQL, its own sandbox scheduler and memory and filesystem checkpoints on gVisor or Kata. A coding Task's state is its checkout, uncommitted changes, artifacts and CLI session, so codemowers/agent keeps it on a PVC and runs workers as ordinary Pods confined by your cluster: Resume starts a new Pod on the same files, with no extra database or scheduler.

Install

Requirements:

• Kubernetes 1.36 or later, with a container runtime that supports image volumes and Pod user namespaces.
• The RabbitMQ Cluster and Messaging Topology operators, and Passmower or another provider of the OIDCClient resource.
• An Ingress controller, DNS and TLS for the dashboard and webhook hosts.
• Admission policies that confine worker Pods, as described in the security model.

The chart does not install CRDs. Apply them from a checkout of the release tag, then install the chart:

kubectl apply --server-side -f crd/
helm install agent oci://ghcr.io/codemowers/charts/agent \
  --namespace codemowers-agent-system --create-namespace \
  --set operator.host=dashboard.example.com \
  --set operator.apiResource=https://dashboard.example.com/api \
  --set webhook.host=webhook.example.com \
  --set webhook.publicOrigin=https://webhook.example.com

The hosts default to the short names dashboard and webhook, for clusters whose admission expands them. operator.apiResource must always be the full dashboard URL plus /api, or sign-in fails with invalid_target. See chart/values.yaml for the other settings.

Next, create the credential Secrets and apply Agents, RBAC, Tasks and receivers from examples/. The dashboard shows copyable kubectl commands and the webhook and device-agent setup for each Agent.

Workspaces and builds

Workers run Ubuntu 26.04 with Node.js, Git, ripgrep, Python, build tools, gh, tea, Helm, Skaffold and the Claude Code and Codex CLIs pinned in runner/package.json. Downloaded tools are verified against runner/checksums.sha256. kubectl and Go are mounted from their upstream images. Install Python packages into a virtual environment in the workspace.

A Task's workspace starts as a Git repository with its specification under specs/. Its PVCs come from the Agent's volumeClaimTemplates; without them Tasks are stateless. Workers build images with Kaniko through Skaffold, using their own ServiceAccount's kubeconfig.

Development

Copy skaffold.env.example to skaffold.env, fill in your cluster's values, apply the CRDs and run Skaffold:

kubectl apply --server-side -f crd/
skaffold dev

Roadmap

• Per-Task broker access: confine each worker to its own Task's queue. Today its token reaches every queue of its ServiceAccount.
• Reopening and key claiming: reopen a completed Task on a later event or user action, and let a worker add routing keys to its own Task.
• Slack: link threads to existing Tasks, including alerts Alertmanager posted to Slack, and copy the first reply there.
• Gitea: test assignment-based Task creation end to end, and discover where repositories can be created.
• Mail: search, and tie sent messages to their Task so replies return to it.
• Worker placement: run workers outside their Agent's namespace, so a hosted tenant's worker memory counts against its own quota.
• Local models: run llama.cpp with reliable tool calls, streaming and session resumption, for the kube-admin profile.
• Separate repositories: these components could run without the operator:
  • mailbox/: an IMAP and SMTP MCP server, with a pluggable bearer-token check in place of TokenReview.
  • webhook/: a signed-webhook-to-AMQP gateway with CEL protocols loaded from files instead of custom resources.
  • operator/devicebridge/: a shell and file MCP server for device-agent.
  • runner/ and src/: an image wrapping the Claude and Codex CLIs with session persistence.

Contributing

Bug reports and pull requests are welcome. Target Kubernetes 1.36, include the tests you ran, and update AGENTS.md when behavior or architecture changes.