Skip to content

Expert plugins and MCP planned

Status: planned, no release number yet

v0.4 opened the hosted platform to declarative specs and built-in models. v0.5 added workflows, governance, the explorers and budget optimisation in the sandbox (what's new). The plan before v0.5 put this work in v0.6 (M4). The releases v0.6 to v0.15 gave other features: the spec builder, scientific models, tree datasets (M4), the science gallery (M5), run tracking, the MMM pack, causal models, run comparison, image data and deep learning. Thus the items on this page are not done yet. With them, allow-listed experts can upload their own model plugins. AI assistants and agents also get a curated MCP server. Simulation-based calibration becomes a single parent run. Sweeps are workflows with a search.grid node (v0.14 retired the Sweep document). Public code upload for all users comes after these items, with its own controls (see the roadmap).

Plugin upload for allow-listed experts

Now, a data scientist can run their own PyMCModel or SklearnModel only on a deployment that they control. They install it into the environment of the runner. This work makes it a product feature of a shared deployment. It is available to tenants that an operator put on the allow list:

flowchart LR
    A["mapi plugin push<br/>(your package)"] --> B["Overlay build<br/>source + pre-resolved wheels,<br/>pinned by hash"]
    B --> C["Describe run<br/>catalog entry, schema sanitizer,<br/>contract kit report"]
    C --> D{{"Kit passes?"}}
    D -->|yes| E["Runnable in your project<br/>on the pinned base image"]
    D -->|no| F["Readable report<br/>of what to fix"]
  • Overlays, not images. The platform builds a plugin into a hash-pinned overlay on top of one base-image version. The overlay has the source and wheels that were resolved before, including native wheels from a curated mirror. Runs never install software.
  • The contract kit runs on upload. The kit does the same checks as mapi plugin test on your computer. These checks decide if a version becomes runnable. The platform keeps the report with the version, next to its SBOM.
  • Schema sanitizer. The platform sanitizes the settings schemas of a plugin before the web app makes forms from them.
  • Dev overlays for fast iteration. The target is less than five minutes from push to a runnable version.
  • A vetted tier. We will port the sibling marketing-mix and HMM models as plugins. They will show that real models agree with the contract.

An MCP server for agents

The MCP server gives a curated set of tools to AI assistants and agents. For example, the tools list and describe models, score an endpoint, read a run and get a plot. The server returns each string from a user or a plugin as clearly delimited untrusted text. Thus a model name cannot put hidden instructions into an agent. Tokens and scopes work the same as for other clients.

Sweeps, SBC and what is already delivered

  • optimize runs (delivered in v0.5). The sandbox optimises budgets next to the model with the planning module of mmm-framework. It uses the response curves of the posterior (Optimise a budget).
  • Backtests (delivered in v0.5) run as workflow nodes on rolling-origin and expanding-window plans (cookbook).
  • Sweeps (delivered). The search.grid workflow node runs each point of a grid and logs the scores of each point as metrics. v0.14 retired the Sweep document: the server does not accept new ones. To compare sweeps, use the compare page.
  • SBC as one parent run that fans out and aggregates the results, for example the simulation-based calibration of a new plugin. (The governance of v0.5 has a parametric-bootstrap version, recovery.simulated.)
  • Bayesian checks in a workflow (delivered in v0.9). The recipe bayes_workflow_audit does a convergence gate, a posterior predictive check, a prior sensitivity check and a recovery check in one workflow run. Since v0.9, recovery.simulated simulates each dataset of a Bayesian model with predictive draws from its posterior predictive. sci_spatial_ladder fits its best model again with wider and narrower priors (Bayesian models in a workflow). These checks run as nodes of the workflow, not as one parent run.

Exit criteria

  • Three sibling models are packaged as plugins and pass the contract kit.
  • A dev overlay is pushed and becomes runnable in less than five minutes.
  • An SBC of a new plugin runs as a single parent.

We cut a release when it passes its exit criteria. Read the changelog.