modelrun

modelrun

modelrun is a small workbench for Replicate. It follows the shape of the service: models live under an owner and a name, each model has versions with their own input schemas, collections gather models worth browsing, and a prediction is one run of a version — or of an official model — with a free-form input object. From that shape modelrun lists and reads models, walks collections and version history, starts predictions, watches them, cancels them, and reports the account the token belongs to. One binary serves a command-line interface and a stdio MCP server.

A Replicate API token is required. Put it in the REPLICATE_API_TOKEN environment variable; it is sent to the API as an Authorization: Bearer header and never appears in output. When the variable is missing, every call fails with an error that names it, before any request is made.

Replicate documents no search endpoint, so modelrun offers none: discovery is list_models and list_collections, narrowed by reading the model you care about with get_model.

Install

npx @reneturcios/modelrun models

Or run the container image:

docker run -i -e REPLICATE_API_TOKEN ghcr.io/cloudcorprecords/modelrun

The container defaults to the MCP server (see below). Pass CLI arguments to use it as a CLI instead.

CLI

export REPLICATE_API_TOKEN=<redacted>

modelrun models
modelrun model meta meta-llama-3-70b-instruct
modelrun collections
modelrun collection text-to-image
modelrun versions stability-ai sdxl
modelrun version stability-ai sdxl db21e45d3f7023abc2a46ee38c23973f6dce16bb082a930b0c49861f96d1e5bf
modelrun predict --model meta/meta-llama-3-70b-instruct '{"prompt":"Explain tides in one sentence."}'
modelrun predict --version db21e45d3f7023abc2a46ee38c23973f6dce16bb082a930b0c49861f96d1e5bf '{"prompt":"a lighthouse at dawn"}'
modelrun prediction zt3k9examplepredictionid
modelrun predictions
modelrun cancel zt3k9examplepredictionid
modelrun account

modelrun --version

Each command prints the upstream JSON payload. Exit codes: 0 on success, 1 when a call fails (including a missing REPLICATE_API_TOKEN or an API error), 2 for usage errors.

MCP server

One bin serves both faces. The mcp subcommand starts a stdio MCP server:

npx -y @reneturcios/modelrun mcp

MCP client configuration:

{
  "mcpServers": {
    "modelrun": {
      "command": "npx",
      "args": ["-y", "@reneturcios/modelrun", "mcp"],
      "env": { "REPLICATE_API_TOKEN": "r8_..." }
    }
  }
}

Tools

Tool	Input	What it returns
list_models	cursor (optional)	One page of models.
get_model	owner, name (required)	The model, with its description and latest version summary.
list_collections	(none)	The curated collections.
get_collection	slug (required)	The collection and the models it contains.
list_model_versions	owner, name (required)	The model's versions.
get_model_version	owner, name, versionId (required)	The version, with its input/output schema.
create_prediction	input (required), exactly one of version / model, webhook (optional)	The created prediction.
get_prediction	predictionId (required)	The prediction, with status and output.
list_predictions	cursor (optional)	One page of the account's predictions.
cancel_prediction	predictionId (required)	The canceled prediction.
get_account	(none)	The account that owns the token.

All calls use the Replicate API under https://api.replicate.com/v1. create_prediction passes the input object through unchanged; for community models pass a version ID, and for official models pass model as owner/name — exactly one of the two, which the tool enforces before any request.

Development

npm install
npm run build   # tsc -> dist/
npm test        # node test runner over test/*.test.ts, no network (fake fetch fixtures)
npm run dev     # run the CLI from source via tsx

License

MIT — see LICENSE.