An agent is most useful in healthcare interoperability when it can inspect a
bounded task and produce evidence the operator can review. It is least useful
when a plausible answer quietly replaces the source material or the human
decision.
Divide the work deliberately
Keep the evidence local
Pidgeon’s local workflows can run with no cloud message-content egress. Keep
message content and detailed technical evidence on the controlled machine, and
pass only the minimum bounded context to an agent or external service that the
team has approved.
“Local-first” describes the workflow boundary, not a universal promise that no
Pidgeon surface ever uses a network. Account, update, licensing, and explicitly
enabled provider features have their own documented boundaries.
Make every finding inspectable
An agent-supported finding should carry:
- the input or stable fixture identifier;
- the exact tool and version;
- the command or operation;
- the relevant rule or comparison;
- the observed output;
- uncertainty or missing context; and
- the human disposition.
If the evidence cannot be inspected, the result is a suggestion, not a
technical conclusion.
The public pidgeon CLI provides a scriptable local interface. The governed
@pidgeonhealth/pidgeon-mcp adapter exposes community capabilities to a
compatible MCP client while keeping the underlying operation in the local CLI
or an explicitly selected desktop Bridge.
Start with the public packages quickstart
and use synthetic data for the first controlled run.
Preserve the human gate
Do not treat these tools as clinical decision support, an autonomous
remediation system, or a substitute for implementation ownership. The agent
can accelerate preparation, execution, and explanation. The accountable human
still defines the boundary and signs off on what happens next. 