goalMet outcome, not just a video link.
How it works
1
Brief
You send an instruction (
"Sign in and create a new project called Acme") plus a target — either a public URL, an environment with a saved session, or a tunnel URL for local targets.2
Sign in, automatically
If an environment is given, Director restores your saved session so it starts already signed in — no pausing to ask for a login.
3
The agent drives
Director carries out the instruction step by step while recording the session. Typed characters are never logged — only a
keyCategory — so credentials never leak into transcripts.4
Render
The raw recording is composited through the Screenforge render engine: automatic cursor-tracking zoom, smoothed cursor motion, captions, and fades — the same engine behind the Screenforge desktop recorder.
5
Verify & ship
Director checks whether the stated goal was actually met, then uploads the final video and returns a signed URL alongside a structured outcome.
Three ways in
MCP
The primary interface — connect an MCP-capable agent (Claude Code, Cursor, or any MCP client) and ask it to create and check on demos in plain language.
CLI
director — used to sign in, connect your agent, and capture authenticated sessions for private or staging environments.Dashboard
Sign in to manage API keys and billing, and to browse the demos your agent has created.
Core concepts
- Demos — one instruction-to-video job. Poll or wait for a structured outcome (
goalMet,stepsCompleted,whatWasShown,confidence) plus avideoUrl. - Environments — a saved target app (base URL + auth). Director can restore a captured browser session or fill a login form before recording, so the agent starts already authenticated. See Environments & Sessions.
- Structured failures — every failure returns a
code, a humanmessage, and aremediationfield so a calling agent can react programmatically instead of parsing English prose.