Kling 4.0 is out for everyone, and your agent can drive it through MCP or a CLI

Kling announced on Oct 15 (9 AM MT) that Kling 4.0 is open to all users, with a redesigned creation interface and access through MCP and a command-line tool. Two items are still marked 'coming soon': 10-bit HDR output at 4K and 1080p, and an API for enterprise customers and developers. Kling's Sep 30 launch post lists the rest: clips up to 30 seconds in one generation, up to 15 reference assets per generation (up to 10 images, 5 videos totalling 30 seconds, and 7 subjects, and a subject can carry a specified voice), up to 10 keyframe images, prompts up to 8,000 tokens, stereo audio with better lip-sync, 21:9, and targeted edits that change one thing (an expression, a camera move, a background) and leave the rest alone. A cheaper, faster Flash tier also exists. Extending a video to 2 minutes is listed as coming soon.

The agent route has two doors. MCP: for international accounts the endpoint is https://kling.ai/mcp. In Claude you add it as a custom connector and sign in with Kling. CLI: npm i -g @klingai/cli-global (Node 18+; China-site accounts use @klingai/cli-cn), then kling login. The CLI is a thin client over the same MCP tools, so command names match the tool names: text_to_video, image_to_video, omni_ref_video (mix images, video, audio and saved Elements), motion_control, element_create for reusable characters, and query_tasks to poll. Output is JSON by default and it never stops to ask questions, so an agent can script it. Run kling who_am_i first. It lists the models and parameters your account can actually use. We haven't checked what model names 4.0 shows up under.

Kling's own guide to batch work with MCP is worth reading even if you never use Kling. Split the brief into what stays fixed (framing, light, pacing) and what changes (character, wardrobe, emotion). Approve the character stills before you animate anything. Use one prompt pattern for every shot: starting state, a visible change, one small action, a camera move, an ending. Test a few clips, and when the same flaw shows up in several, fix the shared rule instead of patching each prompt. Generation is task-based, so the agent submits jobs and polls for them. Every generation spends credits. kling account shows your balance.

For a quality read, TechHalla had early access and made a 90s-style live-action trailer based on the arcade game Cadillacs & Dinosaurs. It's a fan piece, so treat it as a test, not a template for client work. His thread makes three claims, each with the prompt in the attached image. One prompt produced a 7-second fight with 5 cuts. One prompt produced three camera setups (follow her to the car, cut inside, go wide as it pulls out). And Omni Reference with 6 references (the lab set, four heroes, a gang) kept everyone on-model across cuts. He also said he had to change his usual prompting style a bit. Not every reply was convinced about the acting.

Our take: the model is one more strong entry for the picker, and we've added it with Kling's own specs. The bigger news for ReelStack readers is that a top-tier generator now has an official CLI with JSON output. That makes it a building block for Claude Code pipelines next to Higgsfield's MCP. We haven't run it yet. Note that TechHalla's follow-up post includes a Kling invite link. We link his thread, not the invite.