MGPT PlatformContinual MI model platform

Introduction

MGPT is the Continual MI model platform for efficient model execution, masking experiments, local-sized model workflows, and provider-backed model access. The public API surface is intentionally OpenAI-compatible so clients can start with ordinary chat requests while MGPT evolves the underlying execution strategy.

The public base URL is https://mgpt.continualmi.com/api/mgpt/. Local development uses http://127.0.0.1:8090/api/mgpt/ when calling the backend service directly.

  • POST /api/mgpt/chat/completions — text generation through the public MGPT model gateway.
  • POST /api/mgpt/images — image generation for platform and MDL media workflows.
  • GET /api/mgpt/logs — request logs and diagnostics for API traffic.

Chat Completions

Text requests use the familiar Chat Completions shape: choose a model, send messages, optionally stream, and receive an OpenAI-compatible response. MGPT handles model alias resolution, provider routing, monitoring, and usage settlement behind that boundary.

POST /api/mgpt/chat/completions
{
  "model": "deepseek-v4-pro",
  "messages": [
    { "role": "user", "content": "Say hello in one sentence." }
  ]
}

Models

Current public model ids are listed on the Models page. MGPT Platform models use clean product-facing ids while provider-specific routing stays behind the platform gateway.

Compute efficiency

MGPT is oriented around making capable model behavior cheaper and easier to run: smaller local-sized models, prompt/cache discipline, provider adapters, and training data loops that improve behavior without assuming ever-larger prompts are the product.

The long-term target is efficient, controllable intelligence: route requests through the right model, preserve useful diagnostics, and train or distill toward models that can do more with less compute.

Masking

Masking is MGPT's research layer for selective context visibility, cache reuse, and controlled model memory experiments. Public API callers do not need to manage mask commands directly; MGPT keeps those mechanics server-side while the platform exposes ordinary model endpoints.

Authentication

Every endpoint authenticates with a bearer token issued from the MGPT API Keys page. Keys are shown once at creation and stored hashed under the Continual account owner.

Authorization: Bearer cmi_live_<public>.<secret>
Content-Type: application/json

Quickstart

export CONTINUAL_API_KEY="cmi_live_..."

curl https://mgpt.continualmi.com/api/mgpt/chat/completions \
  -H "Authorization: Bearer $CONTINUAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "deepseek-v4-pro",
    "messages": [
      { "role": "user", "content": "Help me inspect this CC1101 register dump." }
    ]
  }'

Diagnostics

MGPT records request status, duration, public model id, provider timing, usage, and error detail for API traffic. Use Logs for request inspection and debugging.

Errors

Errors return JSON with a detail field and an HTTP status code.

{ "detail": "Unsupported model: example-model" }
  • 400 — validation error or unsupported model.
  • 401 — missing or invalid bearer token.
  • 403 — token lacks the required scope.
  • 5xx — provider, routing, or storage failure.