NOT PRODUCTION READY!!
Kubernetes-native runtime for AI coding agents. Submit a prompt + git repo; the platform spins up an isolated workspace pod, runs the agent, and streams results back over ConnectRPC.
The core abstraction is the AgentRun CRD. Everything else — scheduling, LLM routing, approval gates, PR creation — is built around it.
Grab the uncworks CLI from GitHub Releases, then:
uncworks setup # picks a local kube context, installs the Helm chart
uncworks open # port-forward + open the web UICluster requirements: any local Kubernetes (Docker Desktop, OrbStack, k3d, kind). 2 CPU / 2 GiB is the floor; 4/4 is the recommendation.
See docs/getting-started.md for the full path including remote clusters and the TUI.
Runs list — one row per AgentRun, filterable by stage, mode, and approval gate.
Logs — live stream of the agent's actions: prompts, tool calls, file reads, and bash output.
Files — workspace tree inside the agent pod, scoped to /workspace.
Shell — attach to the running pod for ad-hoc inspection.
Traces — span timeline for the workflow, agent thoughts, and tool invocations.
graph LR
User(("User"))
Cloud["OpenRouter / cloud LLMs"]
subgraph K8s["Kubernetes cluster"]
direction TB
subgraph CP["Control plane"]
Web["Web UI"]
API["ConnectRPC API"]
Ctrl["Controller"]
TW["Temporal worker"]
end
subgraph Deps["Deps"]
Temporal["Temporal"]
LiteLLM["LiteLLM"]
Ollama["Ollama"]
Soft["Soft-Serve"]
end
subgraph DP["Agent pod (1 per run)"]
Init["init: hydrate"]
Agent["agent (holds workspace)"]
Sidecar["sidecar: pi-coding-agent"]
end
PVC[("/workspace PVC")]
end
User --> Web --> API --> Temporal --> TW
Ctrl --> API
TW -->|creates| DP
Sidecar --> Agent
Agent --> LiteLLM --> Ollama
LiteLLM --> Cloud
PVC -.- Init
PVC -.- Agent
PVC -.- Sidecar
A run is one Temporal workflow driving one pod. The sidecar fronts pi-coding-agent; the agent reads and writes inside /workspace. Approval gates (LLM judge, human approval) run inside the workflow before a run is marked Succeeded — hybrid (judge + human) is the default.
Spec-driven mode runs three stages with feedback:
sequenceDiagram
actor U as User
participant W as Workflow
participant M as Manage agent
participant I as Implement agent
U->>W: prompt + repo
W->>M: PLAN — write OpenSpec change
W->>I: EXECUTE — write code against spec
W->>M: VERIFY — task gate + spec validate + LLM judge
alt verify fails
W->>I: retry with failure report
end
W->>U: PR opened (autoPush+autoPR)
Single mode skips Plan/Verify — agent runs once against the prompt.
| Where | What |
|---|---|
cmd/{apiserver,controller,worker,uncworks} |
Control plane + CLI |
cmd/sidecar, cmd/hydration |
Pod-side binaries |
internal/server |
ConnectRPC + REST handlers |
internal/temporal |
Workflow, activities, approval gates, LLM judge |
internal/controller |
AgentRun and Project reconcilers |
extensions/aot-determinism.ts |
pi extension loaded into every agent run |
web/ |
React dashboard |
proto/, gen/ |
Service definitions and generated code |
deploy/helm/aot/ |
Helm chart |
devbox shell # tooling
task install
task cluster:setup # one-time: Colima + k3s + Helm install
task dev:deploy # rebuild images into k8s.io and rollout
task dev:web # Vite dev server
task test # Go + web + extension
task proto:gen # regenerate after .proto changestask --list is the rest of the surface. See CONTRIBUTING.md.
- docs/getting-started.md
- docs/architecture/overview.md
- docs/guides/spec-driven.md — Plan/Execute/Verify
- docs/reference/api.md, docs/reference/crd.md
Apache License 2.0 — see LICENSE. Contributions are welcome under the same terms; see CONTRIBUTING.md.