Agent-readable docs index: /site/docs/llms.txt. Full docs in one file: /site/docs/llms-full.txt. Download /site/docs/docs.zip to grep all markdown files locally.
AgentTwin 0.2.2 · discovery and hosted workflow execution · Node.js 22.12+

Safety and isolation

Discovery scans project signals and does not execute application workflows. Review candidate findings, required integrations and missing inputs before preparation.

Execution boundary

Application workflows execute in fresh hosted E2B sandboxes, not local containers. The runtime constructs test credentials and an explicit application environment, runs the application as a non-root user, denies new external guest network connections and checks configured final state. Sandbox deletion is requested after execution with a finite TTL backup.
Custom HTTP collections and the supported Slack subset never forward to production. No provider-native accounts or production OAuth sessions are provisioned or copied. Browser/macOS application execution is not supported by this backend.

Credentials and uploads

The operator’s E2B key and administrative key stay on the control plane. Users receive individual revocable tokens for their own jobs and discovery. The dashboard never saves account tokens in local storage; disconnect and reload clear tab memory.
Discovery sync whitelists reviewed metadata and omits source bodies and credential values. Workflow execution separately uploads explicit reviewed files. Path filtering and generated-secret redaction cannot guarantee that a source file or log contains no sensitive content; review files, fixtures and fix-task evidence before sharing.

Evidence and operating limits

A passing receipt verifies configured assertions against supported behavior; it is not proof of full production equivalence. Unknown operations and unactivated requested faults are reported as incomplete.
The API requires a single replica and persistent storage. Retention, quotas and sandbox TTLs provide bounds. Distributed scheduling, deletion reconciliation, billing and complete monitoring remain future work.
The separate prototype uses fictional data and scripted outcomes. Its simulated diagnosis is not evidence from a sandbox run.