How Trevra works.
Claude Code and Codex call typed modules to do the work. External actions stop at an approval gate, and every run is recorded.
1. Point an agent at a Trevra workspace
Claude Code or Codex connects to a Trevra workspace and calls typed modules to do go-to-market work: research, drafting, and outreach.
2. The module declares its shape up front
Each module is versioned on GitHub with a typed input, a typed output, a declared side-effect class, and whether it needs approval. Installing one never grants it permission to write externally on its own.
3. External actions stop at the approval gate
Sending a message or taking any other consequential action is held for your approval by default. You review the exact payload before anything leaves the workspace.
4. The approved payload is hashed and executed
What you approve is cryptographically hashed before execution, so a modified payload is rejected. The run — its inputs, outputs, and evidence — is recorded.
Modules are the unit of work
A module is a small, typed, versioned unit of go-to-market work, not a freeform prompt. The catalog is public and synced from GitHub, and every module publishes its input and output schemas alongside its side-effect class.
Approval is the default
External actions require approval unless you write a narrow standing instruction with explicit ceilings. Approval integrity comes from the payload hash, not from trust.
Self-hostable on your own PostgreSQL
Trevra is open source. Run the module runner on your own PostgreSQL and keep the ledger, evidence, and configuration in infrastructure you control.