SpacePilot Developer Reference

The definitive technical reference for the SpacePilot high-performance video engine, local hardware bridging, and FastMCP agentic integration.

Local Compute Probe Active

Polling localhost:8080/api/compute

Detected VRAM

-- GB

Active GPU

Scanning...

Recommended Engine

1. Quickstart

Zero-configuration local setup. We rely on Doppler for secrets management. Bypass the $0.30 SaaS markup and run your first $0.00 local inference in under 60 seconds.

# Install CLI and pull config
npm i -g @spacepilot/cli
doppler setup -p spacepilot -c dev

# Run local hardware diagnostics and start daemon
doppler run -- spacepilot daemon start --port 8080

2. Platform Architecture

SpacePilot utilizes a 4-Layer Composable Stack connecting your local editor via FastMCP to either a local Metal/CUDA runtime or a dynamically provisioned Spot Cloud instance based on the Headroom formula.

Headroom Formula: H = VRAM_Total - (OS_Reserve + Model_Weights + KV_Cache + KV_Grad)

3. Model Zoo & DiT Engines

Our tensor compilation targets the optimal precision for each underlying DiT architecture. Below is the engine capability matrix.

4. In-Process Audio Sidechain

Our pipeline features zero-latency in-memory audio sidechaining. We utilize an exact -16 LUFS integrated target, automatically ducking ambient tracks underneath voice (e.g., Kokoro TTS) using a fast lookahead compressor.

5. Agentic FastMCP Reference

Expose SpacePilot capabilities to your agent via FastMCP. The server provides robust schema validation for rendering DocIR 2.0 specs.

{
  "name": "render_video",
  "description": "Submits a DocIR 2.0 payload to the rendering queue.",
  "inputSchema": {
    "type": "object",
    "properties": {
      "prompt": { "type": "string" },
      "engine": { "type": "string", "enum": ["ltx-2.5", "hunyuan"] },
      "aspect_ratio": { "type": "string", "default": "16:9" }
    },
    "required": ["prompt"]
  }
}

from fastmcp import FastMCP

mcp = FastMCP("SpacePilot")

@mcp.tool()
def render_video(prompt: str, engine: str = "ltx-2.5", aspect_ratio: str = "16:9") -> str:
    """Submits a DocIR 2.0 payload to the rendering queue."""
    # Internal routing logic
    return f"Job queued for {engine} at {aspect_ratio}"

6. REST API & Auth Matrix

Direct HTTP access requires the X-SpacePilot-Token header. API interactions are strictly typed.

Endpoint	Method	Description
`/api/compute/local-profile`	GET	Returns host hardware constraints.
`/api/render/submit`	POST	Accepts DocIR 2.0 payload.
`/api/render/status/:id`	GET	Polling endpoint for job completion.

7. Economics

Work runs on the machine you are sitting at, or on one AWS g6e.2xlarge spot instance you launch by hand with spacepilot launch and terminate yourself. There is no automatic overflow router: nothing decides on your behalf that a request exceeds local headroom and sends it to a cloud. The SkyPilot spot-mesh arbitrage this section used to describe was never built, and its API surface has been removed.