The platform
OpenShell runs the agents. Your infrastructure runs OpenShell.
OpenShell is an open standard for operating agents inside an existing engineering environment. It sits between the agents and everything they are allowed to touch: repositories, build systems, issue trackers, and internal services. Access is granted explicitly, policy is enforced at the boundary, and every action is recorded.
What it does
Scoped access
Every agent runs with an explicit grant: which repositories it may read, which branches it may write, which services it may call. Nothing is implicit, and nothing is inherited from a human account.
Policy enforcement
The controls your organisation already wrote — review requirements, protected branches, secret handling, change windows — are enforced at the boundary rather than requested of the agent in a prompt.
Audit trail
Each action is recorded with the prompt, the tool call, the diff, and the identity it ran under. Security and compliance review agent work the same way they review human work.
Native integration
Agents connect to the source control, CI, issue tracker, and internal services already in use. There is no parallel workflow to adopt and no second system of record.
Portable skills
Encoded workflows are stored as code in your repositories, not as configuration inside a vendor product. They survive a change of model, of vendor, and of us.
Model-agnostic
Orchestration is separated from the model. Open-source agents such as Hermes or OpenClaw run alongside commercial models, and either can be replaced without rewriting the skills above them.
What runs where
In your environment
- The OpenShell control plane and its policy configuration
- The agents, their tool calls, and their working state
- Source code, build artefacts, logs, and audit records
- The skills library and the context layer built around it
Outside it
- Model inference, where you choose a hosted model
- Nothing else by default
Where a hosted model is not acceptable for a given repository or dataset, the same skills run against a model deployed in your own environment.
Why it stays in your infrastructure
The value of an agent deployment accumulates in context rather than in the model: which systems it may reach, which conventions the codebase follows, which corrections your engineers made in review. Held inside a vendor product, that context is the thing you cannot take with you. Held in your own environment as code, it remains an asset you own and can move.