Skip to main content
A mirror is a digital twin of the backends an AI agent talks to. It is built from the agent’s production traces (plus optional tool code, schemas and specs) and served as a twin your agent can call instead of production. Concepts: a mirror (one agent’s backends), a build (an immutable artifact from a frozen snapshot of the inputs), the active build (what the twin serves), the twin (a running task with sessions), a session (one isolated world your agent calls tools in), fidelity (how closely the twin answers like production, measured by replaying traces), drift (per-tool reasons the twin differs, each with a lever that fixes it), config (versions of knobs and overrides), proposals (drafted repairs you accept or reject). Flow: identity_whoami; mirrors_builds_intake with traces (see mirrors://formats/inputs) or stream traces with a collector to /api/collect; poll mirrors_builds_get; mirrors_get for the headline; runtime_twin_start; runtime_console_open and runtime_console_invoke to call tools, or open sessions over /v1 with an API key; mirrors_builds_drift and improve_proposals_draft to improve; improve_review_list for what agents changed. Names: tools are the service verbs with dots as underscores; the same verbs back the web app, the CLI and this server. This page is also the MCP resource an IDE agent reads; generated by pnpm docs:generate.