← All integrations

Roboflow MCP server

10 operations0 webhooks

Read workspaces, projects, versions and trained models, create projects, upload images, and run model inference on the serverless host with Roboflow.

Try Roboflow with your agent

Connect Roboflow, then paste this into Claude, ChatGPT or Cursor. For ideas from other apps, see the use cases.

Using Corsair, show me what you can do in Roboflow on my account, then run one read-only example and show me the result.

Runs once, when you ask

Roboflow tools and triggers

Tools are what your agent can do in Roboflow through Corsair's MCP server. Triggers are the Roboflow events a routine can start from.

Project

  • CreateWrite

    Create a project of a given type in an annotation group, optionally licensed onto Universe

    project.create

  • DeleteDestructive

    Delete a project and its images, versions and models

    project.delete

  • GetRead-only

    Read a project with its classes, splits and dataset versions, newest first

    project.get

Version

  • DeleteDestructive

    Delete a dataset version

    version.delete

  • GetRead-only

    Read a dataset version with its image count, splits, and the preprocessing and augmentation it applied

    version.get

Account

  • GetRead-only

    Check that the key works; the route answers 200 either way, so `authenticated` reports whether a workspace actually came back

    account.get

Images

  • UploadWrite

    Upload one image to a project by URL or base64; a duplicate is reported as such rather than as a failure

    images.upload

Infer

  • InferWrite

    Run a trained model on an image on the serverless inference host, returning predictions and their count; billed per image

    infer

Models

  • ListRead-only

    List a project's trained models with their task, architecture and metrics; each `url` is the model ID `infer` takes

    models.list

Workspace

  • GetRead-only

    Read a workspace and its projects with image, annotation and version counts

    workspace.get

Connect Roboflow to Claude, ChatGPT or Cursor

  1. 1

    Connect Roboflow in Corsair

    Add your Roboflow API key once in Corsair. Corsair stores the credential encrypted and makes the Roboflow API calls for your agent, so it never goes into a prompt or a config file. Connect Roboflow

    If your link lets the agent add apps, you can skip this: it gives you a sign-in link the first time it needs Roboflow. On a ChatGPT personal plan, connect Roboflow here first.

  2. 2

    Add Corsair's MCP server to your agent

    Every client uses the same address:

    https://mcp.corsair.dev/mcp
    Open Customize → Connectors → Add custom connector, paste the address, then select Connect.

    Codex, Gemini CLI, VS Code, Zed and other clients with OAuth sign-in work the same way. Clients without it, such as GitHub Copilot's coding agent, use a link's own address and an API key. Setup for every client

  3. 3

    Sign in and choose what needs your approval

    Your client opens Corsair to sign in. Pick a link: the apps your agent can use and a mode that decides what it may do without asking. In the default Cautious mode, Roboflow's read and write operations run right away and its 2 destructive operations wait for your approval. Strict mode asks before every write and refuses destructive operations, and Read mode allows reads only.

FAQ

Install corsair and @corsair-dev/roboflow, add the Roboflow plugin, and connect with an API key. Then call methods like corsair.roboflow.api.account.get() from your app. Corsair keeps each tenant isolated.

Yes. Roboflow requires an API key from your Roboflow account or developer dashboard. Create credentials with access that matches what you automate, then paste them into Corsair or connect through managed auth in Hub.

Yes. One Corsair client can hit Roboflow, GitHub, Notion, Stripe, and the rest. Trigger Roboflow from a webhook, push updates to Slack, or chain actions across plugins. Same tenant context across integrations, plain TypeScript.

Teams use Corsair for read workspaces, projects, versions and trained models, create projects, upload images, and run model inference on the serverless host with Roboflow. Common builds include sync jobs, admin tooling, and agents that call Roboflow APIs from your backend.

The SDK is open source. Self-host and you pay Corsair nothing for the integration layer. Hub bills for hosted auth and execution, not per Roboflow API call. Hub has a generous free tier. For hobby projects you can probably run the whole thing on free. Paid plans start to matter with more tenants or heavier traffic.

We maintain @corsair-dev/roboflow in the open repo. When Roboflow ships API changes, upgrade the package. Your call sites stay typed. You should not need to rewrite the integration layer each time.

No. Corsair does not read your Roboflow data or hold individual access credentials. Credentials stay in your database. API calls run on your instance. We do not store Roboflow data on your behalf.

Yes. Add https://mcp.corsair.dev/mcp as a custom connector or MCP server in Claude, ChatGPT, Cursor, VS Code, Claude Code, Codex or any client that supports remote MCP servers with OAuth sign-in, then sign in to Corsair. On ChatGPT personal plans, custom servers currently get read tools only.

Yes. Put Roboflow on a link in Strict mode and every write waits for your approval, while destructive actions are refused. The default Cautious mode asks only before the 2 destructive Roboflow operations. You can change a link's mode at hub.corsair.dev/me.