How to connect Roboflow to Codex

1. 1. Enable Roboflow in Metorial

   In your Metorial portal, enable the Roboflow MCP server and sign in with the account that owns the data. Access stays governed by your team policies.

2. 2. Add the server and sign in

   The Codex CLI, IDE extension and ChatGPT desktop app share this config.

   Copy

   codex mcp add metorial --url <your-magic-mcp-url>
   codex mcp login metorial

3. 3. Check the connection

   Type /mcp in Codex to see Metorial and its Roboflow tools.

• List Annotation Jobs: List all annotation jobs for a project.
• Upload Annotation: Upload an annotation file for a specific image in a project.
• Run Inference: Run predictions on an image using a trained model deployed on the Roboflow hosted inference API.
• Create Annotation Job: Create a new annotation job for a project.
• Get Project: Get detailed information about a specific Roboflow project, including its dataset versions, annotation classes, image splits, and training status.
• Train Model: Start training a model on a specific dataset version using Roboflow Train.
• List Projects: List all projects in a Roboflow workspace.
• Export Dataset: Export a dataset version as a downloadable ZIP in a specified annotation format.
• Get Image: Get detailed information about a specific image in a project, including its annotations, labels, tags, split assignment, and URLs for the original and thumbnail versions.
• Delete Images: Delete one or more images from a Roboflow project.
• Create Project: Create a new computer vision project in the workspace.
• Create Version: Generate a new versioned snapshot of a project's dataset with optional preprocessing and augmentation settings.
• Search Images: Search for images within a project using text prompts, visual similarity to an existing image, or by filtering on tags, classes, and batch assignments.
• Upload Image: Upload an image to a Roboflow project.
• Get Version: Get detailed information about a specific dataset version, including preprocessing and augmentation settings, training results (mAP, precision, recall), and model endpoint details.
• And 6 more on the Roboflow MCP server page.

• Limit which Roboflow tools Codex sees with enabled_tools and disabled_tools on the metorial server in config.toml.
• Set default_tools_approval_mode to auto, prompt, writes or approve. writes prompts only for tools that are not marked read-only.
• Type /mcp in the composer to see the connected servers.
• Override the approval behavior for a single Roboflow tool with tools.<tool>.approval_mode.

Metorial is listed but tools are missing

Codex can connect without credentials and then see nothing. Run codex mcp login metorial to complete the OAuth sign-in.

Metorial times out at startup

Codex waits 10 seconds by default. Raise startup_timeout_sec under [mcp_servers.metorial] in ~/.codex/config.toml.

Long tool calls fail after a minute

The default tool timeout is 60 seconds. Raise tool_timeout_sec for the metorial server in config.toml.

Based on the Codex MCP documentation.