Configuration you can review.
Describe behavior in manifests. Keep changes in version control so the next model, maintainer or agent can understand the intent.
Own your agents.
Agentify your business.
Build agents around your business. Choose their models, shape their capabilities, and bring them to your team with the control and discipline of software engineering.
PyGubernator makes configuration the foundation you own, version and build on. The vision: Kubernetes-inspired simplicity for creating, deploying and operating agents on your terms.
Current source preview · Unified deployment reconciliation is proposedFROM AGENT LOGIC TO OPERATED WORK
Make the whole agent yours.
Sensitive company knowledge needs approved models and clear access boundaries. Combining agent logic, execution environments, persistence, permissions and operations takes substantial integration work. PyGubernator brings these concerns into one product direction, with configuration as the handoff between building and running.
Choices you control.
Bind model endpoints explicitly. The local source preview keeps hosted inference opt-in and does not silently fall back to an external provider. Start with tool-free local chat.
Work you can return to.
Retain conversations and build on durable workers, lifecycle records and recovery primitives in the current source.
COMPANY-CONTROLLED AGENTS · TARGET EXPERIENCE
Your company’s knowledge. Your company’s agent.
An employee asks an internal agent to prepare a supplier review. It uses approved company knowledge, drafts a recommendation, and routes any external action through configured boundaries. Your team retains oversight of access, sources and outcomes.
- 01Choose where it runs.
Approved local, self-hosted or private model endpoints and execution infrastructure.
- 02Define what it can use.
Company-approved knowledge, scoped memory, tools and permissions.
- 03Give employees access.
An internal agent surface behind the organization’s configured identity and access controls.
- 04Keep oversight.
Versioned changes, retention rules and inspectable action and lifecycle evidence.
This integrated private-operation journey is planned. Authentication, network controls, secret handling and retention require explicit host configuration and qualification for your environment.
KUBERNETES-INSPIRED · TARGET STATEDeclare the intent.
Declare the intent.
Operate the lifecycle.
Our ambition is a Kubernetes-inspired control plane for agents: declare how they should run, inspect their state, and reconcile deployments on infrastructure you control.
A kubectl-like CLI would talk to a durable control service. That service would compare desired and observed state, coordinate workers, and retain evidence of changes and failures.
Agents and the infrastructure behind them- 01Define
Agent revision, environment and boundaries
- 02Deploy
Review bindings and apply approved intent
- 03Observe
Inspect state, outcomes and evidence
- 04Reconcile
Recover or change within declared limits
A FOUNDATION TODAY. A CLEAR TARGET.
Start with what is available.
The current source preview is ahead of the public package. Choose the documented path for the version you can access.
Configure, run and retain the work.
- Validated agent manifests and explicit model bindings.
- Local tool-free chat, settings and stored conversations.
- Durable workers and agent lifecycle primitives.
- Logical provisioning plans and host-authorized placement.
These capabilities span different runtime and host paths. Logical placement binds agents within a configured host; infrastructure setup remains with that host.
A coherent deployment lifecycle.
- Unified deployment bindings and desired-state reconciliation.
- A simple bootstrap reusing the existing runtime and Compose.
- Adapters for already configured remote environments.
- Optional persistent computer sessions and human takeover.
Deployment reconciliation and unified bindings are proposed. Persistent workspace/browser sessions are planned. Their availability will depend on the qualified execution profile.
Public release and source preview are different starting points.
View the published package PyPI 0.0.18, checked on 21 October 2026, predates the local preview. Its wheel has no local extra or up command. Request source-preview access for the newer experience.
THREE PRODUCTS. ONE ENGINEERING APPROACH.
Agents, data and actions. With intent.
Configure each product through its supported interfaces. Bring them together where the work needs them; unified deployment bindings are part of the target state.
PyGubernator
Define agent behavior and coordinate durable work.
You are herePyCharter
Make data requirements executable through pipelines and contracts.
Explore PyCharterPyActuator
Bound external actions with configured policy and audit records.
Explore PyActuator