Roboflow Agent Plugin

Agent-ready Roboflow plugin that contains skills and MCP configuration for computer vision workflows: data management, training, evaluation, inference, model selection, Workflows, Universe, plans, and Roboflow platform APIs.

This repository is a plugin-shaped source of truth for AI agents (Claude Code, Codex, Cursor, OpenCode, and others). The canonical skill content lives in skills/; plugin manifests point at those files instead of copying them elsewhere.

Install as a plugin

The repo ships both plugin manifests pointing at the same skill content and MCP config:

• Claude Code: .claude-plugin/plugin.json — plugin name roboflow
• Codex: .codex-plugin/plugin.json — plugin name roboflow

Both manifests load skills from skills/ and bundle the Roboflow MCP server config from .mcp.json.

The bundled connection uses OAuth: sign in to Roboflow when your client prompts you. No API key is needed. Clients that cannot complete OAuth can configure the API-key fallback.

Microsoft Copilot Cowork

The cowork/ directory contains a Microsoft 365 app manifest, validation, and a Python-only package build that downloads a verified public tool catalog. It uses Roboflow's OAuth Dynamic Client Registration flow and produces a v1.28 package ready for tenant sideloading.

./cowork/build.sh

See cowork/README.md for build and installation instructions.

Claude Code

Install from GitHub — no clone required:

claude plugin marketplace add roboflow/computer-vision-skills
claude plugin install roboflow

The first command registers this repo as a marketplace source (run once per machine). The second installs the plugin.

Per-project installation

To install the plugin only for the current project:

claude plugin install roboflow --scope local

Local scope controls where the plugin is installed. The bundled MCP connection still uses OAuth; it does not automatically read a project API key.

Install from a local clone

git clone https://github.com/roboflow/computer-vision-skills
claude plugin marketplace add ./computer-vision-skills
claude plugin install roboflow

For a throwaway test without touching the installed-plugins list:

cd computer-vision-skills
claude --plugin-dir .

Codex

The Codex CLI currently exposes codex plugin marketplace add, upgrade, and remove. It does not expose a direct codex plugin install command or a codex --plugin-dir flow, so add this repo as a marketplace source and install the plugin from the plugin browser.

Install from GitHub:

codex plugin marketplace add roboflow/computer-vision-skills

Restart Codex, then open the plugin browser:

codex /plugins

Choose the Roboflow marketplace source, select the Roboflow plugin, install it, and press Space if it is installed but still disabled.

Local clone workflow

When editing a local clone, register it as a local marketplace source:

git clone https://github.com/roboflow/computer-vision-skills
cd computer-vision-skills
codex plugin marketplace add .

Restart Codex after edits. If the plugin browser still shows stale metadata, remove and re-add the local marketplace:

codex plugin marketplace remove roboflow
codex plugin marketplace add .

If you registered the GitHub marketplace source instead, refresh it with codex plugin marketplace upgrade roboflow.

What the Codex marketplace file does

Codex reads .agents/plugins/marketplace.json, which points source.path at the repo root via ./. Codex resolves source.path relative to the marketplace root, so the plugin manifest in .codex-plugin/plugin.json, the skills in skills/, and the Roboflow MCP server config in .mcp.json are all loaded from this repository.

Codex caches installed plugins under ~/.codex/plugins/cache/, so a running Codex session may not see edits until Codex is restarted or the plugin is reinstalled from the Plugin Directory.

The bundled Codex connection also uses OAuth. Setting ROBOFLOW_API_KEY alone does not configure MCP authentication; API-key connections need an explicit header configuration as described below.

Install standalone skills

If you want the skills without the MCP server bundle — for example, with an agent that doesn't speak the plugin manifest format — install them directly:

npx skills add roboflow/computer-vision-skills

Install a single skill:

npx skills add roboflow/computer-vision-skills --skill roboflow-inference

By default this installs into ./.claude/skills/ for the current project. Pass -g for ~/.claude/skills/ (global).

The npx skills CLI works with any agent that reads SKILL.md files from .claude/skills/ — Claude Code, Cursor, OpenCode, and others. See vercel-labs/skills for the full CLI reference.

Available skills

• roboflow-api-reference: REST API and inference API references
• roboflow-cloud-storage: connecting S3/GCS buckets to mirror images into a workspace
• roboflow-data-management: uploading images, labeling, dataset organization
• roboflow-inference: running inference, workflows, workflow templates
• roboflow-plans-and-pricing: Roboflow plans and credit usage
• roboflow-product-navigation: where features live in the Roboflow product
• roboflow-training-and-evaluation: training models, diagnosing why a model underperforms, and improving accuracy
• roboflow-universe: searching and using Roboflow Universe

MCP and skills

The Roboflow MCP server exposes live tools for projects, images, annotations, versions, models, Workflows, Universe, and feedback. Skills own the expert guidance and workflow playbooks.

That separation keeps the install model simple:

• MCP server: live Roboflow tools and authenticated API access
• Plugin skills: durable product guidance and workflow playbooks
• This repo: canonical source for skill updates and plugin distribution

API-key fallback

Use this only when your client cannot complete OAuth or you need a headless API-key connection. Obtain a private key from Roboflow settings, then configure your client to send it in the x-api-key header to https://mcp.roboflow.com/mcp.

For clients that support environment expansion in MCP JSON configuration:

{
  "mcpServers": {
    "roboflow": {
      "type": "http",
      "url": "https://mcp.roboflow.com/mcp",
      "headers": {
        "x-api-key": "${ROBOFLOW_API_KEY}"
      }
    }
  }
}

Load ROBOFLOW_API_KEY into the environment of the process running the client, using a secret manager or a gitignored .env that you explicitly load. A .env file is not automatically loaded by this plugin. If your client uses a different configuration format or does not expand environment variables, use its supported secret-to-header configuration instead of this JSON example. Configure one Roboflow connection, avoiding duplicate plugin and manual server entries.

When switching an existing OAuth connection to an API key, sign out of that connection or clear its saved OAuth credentials so the client stops sending an Authorization: Bearer header. The server prefers that header over x-api-key when both are present, so adding the API-key header alone may still use the OAuth workspace and permissions.

Never paste a private key into chat or commit it. To return to OAuth, remove the API-key header configuration and reconnect.

Contributing

Skills are markdown. Open a PR with edits or a new folder under skills/. Each new skill must have a SKILL.md at its root with name and description frontmatter.

License

Apache-2.0