AgentCastKit

AgentCastKit is an agent-first screen-demo system for macOS. An AI agent supplies a semantic demo plan through MCP; a local native worker owns ScreenCaptureKit and permission-sensitive work; durable jobs make long recordings and renders resumable and inspectable.

This repository is the first vertical slice. It can:

• expose a local MCP server over stdio;
• inspect capture permissions without prompting;
• request one named permission only with explicit confirmation;
• list recordable displays and windows;
• validate a privacy-aware semantic demo plan;
• start a bounded display/window recording and return a durable job ID;
• install and configure a pinned Cua Driver companion so agents can drive native apps and browsers while capture continues;
• install production-grade AgentCastKit and Cua Driver skills for detected Codex and Claude Code hosts;
• expose a machine-readable free/paid feature policy;
• query a provider-neutral marketplace voice library;
• synthesize metered voiceovers through the activated AgentCastKit service without exposing cloud credentials;
• inspect completed video artifacts and analyze visual activity on device;
• produce an editable, non-destructive keep-list that trims dead time while preserving the raw take;
• render local MP4 edits as durable jobs, including ordered cuts and playback-rate changes;
• persist job state atomically and reconcile interrupted work after restart.

GUI and browser control run through Cua Driver's companion MCP. AgentCastKit 0.4 includes a conservative visual-activity editor and local MP4 renderer. It does not yet persist semantic action telemetry, capture editable separate tracks, analyze speech, or add automatic zooms, captions, music, and narration mixing. Those boundaries are explicit in recorder_describe.

This is the open-source local runtime. AgentCastKit's hosted activation, billing, protected orchestration recipes, licensed media, and provider integrations are maintained separately and are not part of this repository.

Install with npx

On an Apple silicon Mac running macOS 15 or newer:

npx agentcastkit install

The installer verifies the bundled notarized Runner, installs it to ~/Applications, downloads the pinned and checksummed Cua Driver 0.22.2 release, opens both signed apps' permission experiences, installs advanced agent skills, and configures detected Codex and Claude Code clients with both MCP servers. Cua Driver telemetry is disabled by the AgentCastKit installer. Neither local recording nor computer control requires AgentCastKit activation.

Useful follow-up commands:

npx agentcastkit doctor
npx agentcastkit mcp-config

macOS Runner app

Build the native GUI runner, embedded MCP runtime, Developer ID signature, and distributable ZIP:

npm run native:app
open "build/AgentCastKit Runner.app"

After configuring a notarytool Keychain profile named AgentCastKit, produce a fresh signed, notarized, and stapled release with:

npm run native:release

The runner checks and requests Screen Recording, microphone, and camera permissions; activates against the Laravel API while keeping the bearer token in Keychain; and copies a ready-to-paste MCP configuration for both runtimes. Cua Driver has its own stable signed app identity for Accessibility and Screen Recording permission. The same signed AgentCastKit executable presents the GUI when opened normally and implements the native JSON protocol when an MCP host invokes it with command-line arguments. Its cloud broker performs only allow-listed AgentCastKit API operations, so neither the activation credential nor an underlying provider key appears in an MCP tool response.

Run it

Requirements: macOS 15+, Xcode command-line tools, and Node.js 22+.

npm install
npm run native:build
npm run build
node dist/src/server.js

The last command is an MCP stdio process, so it waits for JSON-RPC on stdin rather than presenting a terminal UI.

For a development client configuration:

{
  "mcpServers": {
    "agentcastkit": {
      "command": "node",
      "args": ["/Users/obaidahmed/Development/agentcastkit/dist/src/server.js"],
      "env": {
        "AGENTCASTKIT_CAPTURE_BIN": "/Users/obaidahmed/Development/agentcastkit/native/.build/debug/agentcastkit-capture"
      }
    }
  }
}

Call recorder_describe, product_features, permissions_status, and then sources_list. Rehearse the flow through the Cua Driver MCP before starting capture. After a completed take, call artifact_inspect, edit_analyze, edit_plan_get, and render_start; all post-production remains local and preserves the raw source. Call voice_library_list only for an entitled account, then pass one returned voiceId to voiceover_synthesize. Voice previews are opt-in so the default response stays compact. Permission prompts and recordings require explicit user confirmation. Captures default to:

~/Library/Application Support/AgentCastKit/recordings/

Voiceovers default to:

~/Library/Application Support/AgentCastKit/voiceovers/

Rendered edits default to:

~/Library/Application Support/AgentCastKit/renders/

Development

npm run check

The example plan is in examples/browser-demo-plan.json. The architecture and next milestones are in docs/architecture.md.

Open-source boundary

Apache-2.0 covers the npm installer, MCP server and protocol, native macOS capture runner, documentation, and examples in this repository. It does not cover AgentCastKit cloud services, private prompts and production recipes, subscription systems, provider credentials, premium music, or hosted TTS/rendering services.

The free local product includes recording, Cua Driver control, planning and validation, durable jobs, local artifacts, non-destructive editing, and MP4 rendering. Paid services are managed TTS, premium or cloned voices, premium licensed music, brand kits, hosted video, and team collaboration. Local artifacts are never uploaded automatically.

Contributions are welcome; see CONTRIBUTING.md. Please report vulnerabilities through GitHub private vulnerability reporting, not a public issue.