Demos
Watch Orka workflows on Kubernetes, grouped by feature. Each demo links to the documentation for its setup and behavior.
Watch the full playlist on YouTube.
Chat and pull requests
Follow a coding-client request through coding and review Tasks, then verify the Publisher's pull request on GitHub. The client uses a cluster token rather than a model API key.
Read about interactive chat, the Anthropic-compatible API, and workspace publication.
Execution workspaces
Stop an idle worker, resume the same Session with its files intact, and clean up the workspace when the work is done. This demo uses Agent Sandbox workspaces.
Save an independent data-only checkpoint, delete the original workspace, and restore its files into a new workspace. This demo uses Agent Substrate workspaces. It preserves files, not a running process or its memory.
Repository security scanning
Register a test repository, review its threat model and validated findings, then choose a finding to fix and open a remediation pull request.
Read the repository security scanning guide.
Shared AI endpoints
Route each caller to its team's Orka installation through one endpoint, then check that requests for the other team's models and namespace are refused.
Read about shared OpenAI and Anthropic endpoints.
Agent-to-agent requests
Discover an Agent Card, ask an inventory agent for advice through A2A, retry without duplicate work, and continue the conversation in the same Session.
Read about the A2A gateway adapter.
Tool access controls
Expose a stock lookup while keeping the supplier's purchasing route closed. Verify that the lookup reaches the test supplier and no order is created.
Read about Tool configuration and outbound access policies.
Memory
Let an agent propose a note, review it, and explicitly apply it before a fresh agent can use the shared knowledge. Accepting a proposal alone does not publish it.
Read about memory proposals and durable memory.
Human approval
Hold a proposed work order for review, record a person's approval, and verify that exactly one order is created. The result returns to the original waiting Task.
Read the human approval guide.
Provider routing
Compare hosted execution with Vekil routing between local AIKit inference and a hosted model, including fallback behavior, elapsed time, and estimated cost.
Read the provider proxy guide.