Skip to main content

Introduction

The Labellerr MCP Connector enables autonomous multi-step labeling workflows through conversational prompts. Claude chains all 23 tools intelligently: creating datasets, establishing label templates, deploying projects, checking annotations, monitoring live progress, and exporting ground truth.
All workflows below use the Hosted Connector. Ensure you have connected Labellerr in Claude settings: see Overview & Setup.

Full Workflow: Bucket to Export

You can execute a complete end-to-end computer vision pipeline through Claude in a single conversational thread:
Complete Pipeline
1

Step 1: Connect your cloud storage

Prompt
Claude calls connect_cloud_source. You save the bucket connection securely in the Labellerr web app, then tell Claude when finished. (See Connect Cloud Storage).
2

Step 2: Import data as a dataset

Prompt
Claude calls list_connections, retrieves the connection ID, invokes dataset_create, and tracks the job with get_job_status until the dataset is ready.
3

Step 3: Set up the project in one step

Prompt
Claude calls project_setup, creating both the label schema and the annotation project together.
4

Step 4: Load pre-annotations (with dry-run validation)

Prompt
Claude validates file names, label categories, and coordinate boundaries (dry-run check), then calls upload_preannotations (or upload_preannotations_async) and monitors the import.
5

Step 5: Label and review in Labellerr

Annotators, quality auditors, and reviewers complete labeling and review tasks in the Labellerr web interface as normal. You can ask Claude “What is the progress on project Vehicle Detection?” to inspect completion rates via monitor_project_progress.
6

Step 6: Export annotations

Prompt
Claude calls export_annotations, tracks completion with get_job_status, and delivers a signed download link via download_export.
All-in-One Prompt: You can combine these commands into a single request: “Import vehicles/2026-11/ from my S3 connection as ‘Vehicles Nov’, then create an image project ‘Vehicle Detection’ on it with boxes for car, truck, and bus.”

Finding Your Work

Quickly survey your assets across projects, datasets, templates, and storage connections:
Prompt Examples

One-Step Project Setup from Label List

The fastest way to configure a project. Simply specify label names and the dataset; Claude handles the rest.
Prompt
Claude translates this into an atomic project_setup call:
project_setup Payload
  • Output: Returns project_id and template_id. The project is immediately visible in the Labellerr web app with all three label classes configured.
The target dataset must be in ready status and contain files. If you do not yet have a dataset, Claude will prompt you to import one first.

Projects with Advanced Question Types

For projects requiring multi-class classification, radio buttons, dropdown menus, boolean flags, or freeform text inputs, configure the template explicitly:
Prompt
Claude executes the following workflow:
  1. Calls template_create with the structured question schema.
  2. Calls project_create linking the new template_id to Vehicles Nov.
  3. Configures review_rotation_count: 2.

Pre-Annotations & AI Pre-Labeling

Accelerate labeling by seeding projects with model predictions or Claude-assisted pre-labels.

1. Upload COCO Pre-annotations

Prompt
  1. Send payload: upload_preannotations (or upload_preannotations_async for large datasets) sends the COCO JSON and receives a job_id.
  2. Backend processing: Labellerr matches each image entry in the COCO file against dataset filenames.
  3. Status & audit: get_job_status returns succeeded alongside an audit report of any unmatched files or unmapped categories.
Image filenames in the COCO file must match dataset filenames strictly, including case and file extension (e.g., IMG_1.jpg vs IMG_1.JPG). Maximum file size is 25 MB.

2. COCO Dry Run Validation Available

Verify a COCO JSON file against your project before committing annotations:
Prompt
Claude performs a pre-flight dry run:
  • Identifies exact filename matches and warns of case mismatches (frame_01.PNG vs frame_01.png).
  • Flags categories present in the COCO file that are missing from the project template.
  • Detects bounding box coordinates that exceed image dimensions.
  • Presents a summary report for your approval before writing any data.

3. Claude-Assisted Pre-Labeling with Review Available

Let Claude propose initial bounding boxes and preview them for human review:
Prompt
  1. Claude inspects sample images using Claude Vision.
  2. Claude predicts bounding boxes matching the project’s label schema.
  3. Claude presents the proposed boxes visually or as structured coordinates for your inspection.
  4. Upon your confirmation (“Looks good, apply to the project”), Claude uploads the annotations as pre-labels for human annotators to refine.

Live Progress & Analytics Monitoring Available

Monitor project metrics and annotator throughput in real time:
Prompt
Claude queries monitor_project_progress and query_project_statistics:
  • Total files vs completed vs review stages
  • Completion percentage and audit status
  • Distribution across bounding box classes and categories

Google Drive Import Workflow Available

Import image batches directly from shared Google Drive folders:
Prompt
Claude executes:
  1. dataset_create with the Google Drive link (supports up to 2,500 images and 2.5 GB).
  2. Monitors import completion with get_job_status.
  3. Immediately invokes project_setup once the dataset is ready.

Exporting Annotations

Export finished annotations for model training or downstream analytics:
Prompt
Signed export download links expire after a short duration for security. If your link expires, simply ask Claude: “Generate a new download link for export exp_abc123 in project proj_456”.

Resuming Long Tasks Across Sessions

Dataset imports, large pre-annotation uploads, and complex exports run asynchronously on Labellerr servers. If you close your browser or Claude chat, your tasks continue uninterrupted. Every asynchronous operation returns a job_id:
  • Dataset imports: ds:<dataset_id>
  • Pre-annotation uploads: pre:<project_id>:<job_id>
  • Exports: exp:<project_id>:<export_id>
In any future Claude chat, simply ask:
Prompt
Claude calls get_job_status and returns the current status, completion timestamp, or download link.

MCP Tools Reference

Detailed specifications for all 23 tools and parameters.

Connect Cloud Storage

Secure AWS S3, GCS, and Google Drive integrations.

Upload Pre-annotations (SDK)

COCO format standards and Python SDK pre-annotation methods.

Create Exports (SDK)

Explore export formats and metadata structures in detail.