mere.run plugins

This repository is for people who use or extend mere.run production workflows. It contains the official plugin catalog, shared contracts, workflow recipes, and companion command-line interfaces (CLIs).

mere.run owns model installation and inference. Plugins own the work around inference, including planning, orchestration, artifact records, post-processing, provider resources, and cleanup.

• Read the documentation
• Choose a plugin
• View the live catalog
• Read the plugin contract

Install a plugin

To inspect the available plugins, list the catalog:

mere.run plugin list

To preview an installation, omit --yes:

mere.run plugin install mere-image-tools

To install the plugin, confirm the catalog command:

mere.run plugin install mere-image-tools --yes

After installation, check the plugin and inspect its machine-readable manifest:

mere-image-tools doctor
mere-image-tools manifest --json

For task-specific installation and first-run instructions, see the getting-started guide.

Choose a plugin

The catalog contains 18 companion executables:

Task	Plugin
Analyze geospatial data	Geospatial Tools
Index and search faces in a photo library	Face Tools
Index and search a shared drive	Archive Tools
Create visual-effects shot artifacts	VFX Tools
Remove a still-image background	Image Tools
Produce animatic assets	Animatic Tools
Produce a governed short film	Film Studio
Run a live local music performance	Perform
Publish review artifacts	ShotGrid Tools
Run LoRA training on a user-owned pod	RunPod Runner
Compare local agents with Terminal-Bench	Terminal-Bench
Run identity-adapter workflows	Identity Tools
Convert or process documents	Document Tools
Remove sensitive text from media frames	Media Scrub
Prepare image-caption datasets	Dataset Tools
Transcribe and reduce sensitive text	Transcript Tools
Generate images with references or adapters	Image Compose
Run resumable local command batches	Batch Runner

For selection guidance, see Choose a plugin.

For authored local model evaluations, see the ATE-derived tool-decision cases. This source pack includes 240 cases and 60 executable fixture cases; it is not an installable plugin or an official Cohere benchmark.

Understand the safety boundary

Plugins are separate executables. mere.run doesn't load plugin code into the core process.

Every provider plugin must support doctor, plan, run, resume, and cleanup. Before a plugin creates a paid resource, it must write run.json and provide a plan or dry run. Remote providers terminate resources by default unless you select an explicit keep or debug option.

Commands that promise JSON write machine-readable data to stdout. They write diagnostics to stderr. Plugins must not write credentials or secrets to either stream.

For the complete rules, see Provider safety and Plugin security.

Develop a plugin

The repository uses Python packages for plugin executables and VitePress for documentation. The main directories are:

benchmark-recipes/  Pinned external benchmark protocols
benchmarks/         Authored evaluation cases and deterministic fixtures
bundles/            Reviewed inputs for signed macOS bundles
catalog/            The official installation catalog
contracts/          Language-neutral JSON Schemas
docs/               VitePress documentation source
eval-recipes/       Reference evaluation protocols
packages/           Python plugin packages
recipes/            Executable workflow recipes
scripts/            Validation and maintenance commands

For repository structure and contribution steps, see Development guidance and CONTRIBUTING.md.

Validate a change

Before you open a pull request, run the repository gate:

./scripts/check.sh

To validate and build the documentation, run:

corepack pnpm install --frozen-lockfile
corepack pnpm docs:coverage
corepack pnpm docs:build

The repository gate compiles packages, runs unit tests, validates contracts and recipes, and tests installed commands. The Terminal-Bench smoke test remains plan-only and verifies that the plugin doesn't create a Docker runtime.

A successful local build doesn't prove that a package, catalog, or website was published. For release boundaries, see Releasing and deployment.