Put your agent to work on your models.

Features · MCP Server

Connect the MCP client you already use to Prediction Lab on your machine, and let it branch a model, run a fit, and open a review.

Your client

An MCP client connected to pl-mcp takes the prompt to try driver_age as a spline in the frequency model on a branch. It reads the model, asks before its first write and is allowed for this chat, then branches the model, adds driver_age as a natural spline, refits, compares the driver_age effect, and opens a review into Default. Beside the transcript, a chart shows the driver_age effect with the current fit and the spline fit, above the open model review. Replay runs it again and asks before each write until you allow it for the chat or deny it.

Claude Desktop connected to pl-mcp

local · 6 tool calls

Prompt: Try driver_age as a spline in the frequency model on a branch, refit, and open a review if it helps

  1. Read the frequency model

    describe_modelDone

  2. Allow create_branch?

    Allowed for this chat

  3. Branched the model as driver-age-spline

    create_branchDone

  4. Added driver_age as a natural spline

    add_featureDone, allowed for this chat

  5. Refit on the branch

    run_model_fitDone, allowed for this chat

  6. Compared the driver_age effect

    compute_model_effectsDone

  7. Opened a review into Default

    create_reviewDone, allowed for this chat

Run complete. A review is open into Default.

driver_age effect

Model review

Open

driver-age-spline → Default

Setup

A few minutes, outside the app.

pl-mcp is opt-in. It runs in a separate MCP client such as Claude Desktop, Claude Code, Codex, or Cursor, not inside Prediction Lab. It is available on macOS and Windows.

  1. Open the MCP Server page.

    In the desktop app, choose MCP Server from the avatar menu at the bottom of the sidebar. The first time you open it, the app installs the signed pl-mcp binary on your machine.

  2. Add pl-mcp to your MCP client.

    Copy the path from the app's MCP Server page into the snippet shape shown for your client, add it to your client's config, then restart the client so it picks up the new server.

  3. Keep Prediction Lab running.

    pl-mcp talks to the desktop app on your machine. When the app is closed, the server has nothing to connect to.

claude_desktop_config.json

{
  "mcpServers": {
    "prediction-lab": {
      "command": "<path shown on the MCP Server page>"
    }
  }
}

Capabilities

The same work you do in the app.

Your client works in your local workspaces, and its changes land in Prediction Lab like any other. Have your agent work on a branch and open a review when you want a teammate to check it first.

Explore
Browse your workspaces and datasets and read their rows.
Example prompt: “Which of my datasets have a claim_count column?”
Tools:
  • list_datasetslist datasets
  • preview_rowspreview rows
Ingest
Import a local file as a new dataset, or replace an existing dataset's file.
Example prompt: “Import policies_2026.csv from my Downloads folder as a new dataset.”
Tools:
  • ingest_datasetimport a file
  • replace_dataset_filereplace a file
Prepare
Set the target, add features, and filter the rows a model sees.
Example prompt: “Add vehicle_age as a feature and leave out commercial policies.”
Tools:
  • create_targetset the target
  • add_featureadd a feature
  • set_filterfilter rows
Model & fit
Create a model, run a fit, or have the app build a model for you.
Example prompt: “Build an automated frequency model on claim_count and fit it.”
Tools:
  • create_modelcreate a model
  • run_model_fitrun a fit
  • build_automated_modelbuild automatically
Inspect
Read a fit's results and compare how an effect changed.
Example prompt: “How did the driver_age effect change between the last two fits?”
Tools:
  • get_model_fitread a fit
  • compute_model_effectscompare effects
Collaborate
Work on a branch and open a review for a teammate to check.
Example prompt: “Make this change on a branch and open a review into Default.”
Tools:
  • create_branchbranch
  • create_reviewopen a review
  • decide_reviewapprove or close

The server also ships ready-made prompts, such as analyze-dataset. Clients that support MCP prompts show them as slash commands.

Control

You decide what it can change.

Your MCP client talks to the Prediction Lab MCP server, pl-mcp, over stdio, and pl-mcp talks to the Prediction Lab desktop app. All three run on your machine.
Read-only mode
Write tools are on by default. Start pl-mcp with --read-only and your client can read your data but can't change it. Add --allow-compute to turn compute back on under read-only.
{
  "mcpServers": {
    "prediction-lab": {
      "command": "<path shown on the MCP Server page>",
      "args": ["--read-only"]
    }
  }
}
Approval before writes
pl-mcp marks every tool that changes your workspace as destructive, which is the signal MCP clients use to ask before running it.
Local by design
pl-mcp connects only to the app on your machine. There's no token to manage and nothing listens on the network. If your client is a hosted AI, what it reads through pl-mcp goes to that client's provider.
Branch and review
Your agent works on a branch. You merge it. How reviews work

More features

Explore the rest.

Ready to ship better models?

Download Prediction Lab, or book a 15-minute demo with our team.

Download