The team you tuned is the team you deploy.
The runtime of the first trial on your machine is the one that runs on the server.

Only the configuration changes
No redevelopment: the logic written for the trial is not redone.
No dead end: if complexity grows, you change notation without changing foundation.
The boundary follows the team: mounts, budgets and circuit breakers are in the runtime. An autonomous run operates under the budget the team declares. Without a declared budget, neither tokens nor duration are capped. In the other modes, the token and cost cap is the one the host (the application that starts the runtime) sets; there is none by default.
What changes from one environment to another is the configuration: the model used, the mounted folders, the secrets.
Deployment, state, observability, model provider
A team is deployed like software (new tab)
The
orkeonCLI, a container, an installer for Windows (.msi), an apt package for Debian or Ubuntu, an archive for macOS,orkeon-host(the service that hosts teams on a server), or a library in a .NET application.Durable state is opt-in (new tab)
Memory in a durable store and checkpoints are used to resume an interrupted run. The model’s responses can be recorded and replayed; two evaluation reports can be compared.
Runs are observed through OpenTelemetry (new tab)
Traces, metrics and logs go to the collector you name. Each run has its own stream of events.
The model provider is changed through configuration (new tab)
The
--llm-profile <id>option picks, for one run, one of the providers declared underLlm:Profilesin the configuration. The team and its boundary do not change.
The known limits of deployment are detailed on the Security page.
- The container image carries no provenance attestation (new tab): pin it by its digest (
@sha256:…). RetentionDaysis not enforced yet: rotating the audit is up to you.--llm-logkeeps every exchange with the model in plain text: this file must be protected and purged.orkeon-hostis neither a cluster nor a scheduler; it comes from the full archive on the releases page (or from its.msion Windows), not from the apt package.- In a container, shell confinement is turned off.
- The audit does not record the arguments of tool calls.
- What an agent reads is sent to the model. With a cloud provider, this data leaves the company; with a local model, it stays there.
The same team runs with two providers
--llm-profile option differs. The same team runs twice, with the same configuration file.The configuration file
{
"Llm": {
"BaseUrl": "http://localhost:11434/v1", "Model": "qwen2.5:1.5b",
"Profiles": {
"claude": { "BaseUrl": "https://api.anthropic.com/v1", "Model": "claude-sonnet-5", "ApiKeyEnvVar": "ANTHROPIC_API_KEY" },
"deepseek": { "BaseUrl": "https://api.deepseek.com/v1", "Model": "deepseek-v4-flash", "ApiKeyEnvVar": "DEEPSEEK_API_KEY" }
}
}
}$ orkeon run crew.yaml
$ orkeon run crew.yaml --llm-profile claude -v 1LLM profile claude elected as the run's default (--llm-profile)The team from your machine is checked in the container, without a call to the model (installation guide, Container section (new tab)). Exported over OTLP, a run shows each tool call as an execute_tool trace (ADR-011 (new tab)).
Documentation
- The
orkeon-hostservice (new tab)/architecture/service-host - Install with apt (new tab)/guides/install-with-apt
- Replay, evaluation and
--llm-log(new tab)/reference/cli - Known limitations (new tab)/reference/limitations