Public documentation

Roleplay Documentation

Run locally

Install the included CLI, choose a provider and judge, and run a social-engineering pack locally.

Connect the workbench

Create a project, save the API key, configure your LLM provider, and upload safe finding summaries.

Review exploit proof

Inspect findings, failed invariants, evidence, fixes, and regression keys.

Add the CI gate

Run the same scenario or pack in release workflows and catch regressions.

Roleplay Documentation

Roleplay is a paid workbench for social-engineering tests for AI agents.

Use it to test whether an agent can be manipulated into violating a protected boundary, policy, tool-use rule, or user intent. The core product loop is:

Create a Builder or Team workspace.
Create a project, protected agent, and project API key.
Choose attacker and judge providers for real adaptive runs.
Run attack simulations locally or in CI with the included CLI.
Upload safe finding summaries to the workbench.
Review Evidence, assign the fix, and keep the same check in CI.

Who This Is For

Roleplay is built for teams shipping:

customer support agents
billing and refund agents
internal copilots
browser or tool-using agents
agent workflows that read untrusted content

Plans

Builder

Builder is about $49/month for solo founders, consultants, and developers testing one or a few AI agents.

Team

Team is about $199/month for teams and agencies that need shared findings, owners, run history, CI uploads, and regression detection.

The CLI is included as the local execution engine for both plans. Real adaptive runs currently require your own provider key and explicit judge mode; mock mode is only a smoke test.

Privacy Defaults

Roleplay is local-first by default.

Full transcript upload: off by default
Sanitized finding summaries: on by default for Cloud uploads
Redacted snippets: on
Secret redaction: on
Default evidence retention: plan-based in the workbench

Start Here

Documentation Map

For local execution:

For scenario authors:

For workbench users:

For integration and operations: