Discoverable
Configured tools and resources can describe what they do and what input they require.
Documentation / MCP
Tools should be explicit, inspectable, and purpose-bound.
Model Context Protocol gives Rodon a consistent way to connect tools and resources. The protocol is a capability boundary—not automatic permission to use everything a server can reach.
Assignable
Administrators and agent builders decide which capabilities belong to which agent.
Auditable
Tool use can carry user, conversation, agent, operation, and correlation context.
How it works
Trusted development tools remain separate.
A coding MCP server with repository, shell, SSH, or Docker access belongs to a different trust category than a disposable model sandbox.
- Rodon Compute must not route into a trusted Codex MCP server.
- Agent permission to see a tool does not automatically authorize consequential use.
- Secrets belong in server configuration, never agent instructions or browser state.
- Failures should expose a useful normalized error without leaking credentials.