Roboflow MCP Server
The Roboflow MCP Server enables programmatic management of Roboflow datasets and projects from MCP-compatible clients (e.g., Claude Desktop, Cursor). Key capabilities include:
Workspace & Project Management
Retrieve workspace metadata and list all projects
Get detailed project info (classes, image count, splits)
Image Management
Upload single or batch images via URL, local file path, or base64 encoding
Search images by tag, class name, or semantic prompt (with pagination)
Retrieve image metadata or delete images
List upload batches for a project
Image Tagging
Add, remove, or replace tags on images
Annotation Management
Upload annotations in COCO, YOLO, Pascal VOC, CreateML, or Roboflow JSON format
List annotation jobs with optional status filtering
Promote fully-reviewed annotation jobs into the dataset with configurable train/valid/test splits
Dataset Versioning & Export
List, get, create, and delete dataset versions
Initiate async version generation with custom preprocessing and augmentation settings
Poll generation status (queued, generating, ready, failed)
Export versions in formats like COCO, YOLOv5/v8/v11, Pascal VOC, TFRecord, and more
Download exported dataset ZIP files to disk
Safety Controls Destructive operations (delete image, delete version, download to disk, etc.) require explicit confirmation and appropriate server mode.
Exposes the Roboflow API to manage datasets, upload and tag images, push annotations, and inspect projects.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@Roboflow MCP Server list datasets in my workspace"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
mcp-server-roboflow
An MCP server that exposes the Roboflow API to MCP-compatible clients like Claude Code, Claude Desktop, and Cursor.
Use it to manage datasets, upload and tag images, push annotations, and inspect projects without leaving your editor.
Not affiliated with Roboflow Inc. This is a community project.
Status
Early development. The public API and tool names will change until v1.0.0.
See CHANGELOG.md for what shipped in each release.
Related MCP server: claude-bridge-mcp
Install
Requires Python 3.10 or newer and a Roboflow private API key.
Claude Code
claude mcp add roboflow \
--scope user \
--env ROBOFLOW_API_KEY=your_key_here \
-- uvx mcp-server-roboflowRestart Claude Code. Tools appear as mcp__roboflow__*.
From source
git clone https://github.com/MayankD409/Roboflow-MCP-Server.git
cd Roboflow-MCP-Server
uv sync
uv run mcp-server-roboflowConfiguration
Environment variables (copy .env.example to .env for local dev):
Variable | Required | Description |
| yes | Your Roboflow private API key |
| no | Default workspace slug, so tools can omit it |
| no | Override the API base (default |
| no | Override the app host used by session-cookie tools (default |
| no | Browser-session Cookie header for |
| no |
|
Tools
See docs/TOOLS.md for the full list. v0.1 focuses on dataset management: list workspaces and projects, upload images, add and filter by tags, upload annotations.
Contributing
We follow Git Flow and test-driven development. New tools land through PRs to
develop; releases cut from develop into main with a semver tag.
Read CONTRIBUTING.md before opening a PR. Bugs and feature ideas go in issues.
License
Available Tools
11 toolsroboflow_add_image_tagsC
Add one or more tags to an image. Additive; not destructive.
Inputs: tags (required), project (required), image_id (required), dry_run, +1 more. Returns: object
roboflow_add_reviewed_to_datasetA
Add every fully-reviewed annotation job to the dataset.
Moves images from jobs in review with all images approved into
the dataset, split randomly at the given global ratios
(default 70/20/10). Restrict to specific jobs via job_ids.
Destructive: requires confirm='yes' and a server mode of
curate or full. Also requires ROBOFLOW_SESSION_COOKIE
(an app.roboflow.com browser-session cookie) because Roboflow
has no public API for this action.
Inputs: project (required), confirm, dry_run, job_ids, +4 more. Returns: result
roboflow_create_versionB
Kick off an async version generation (quota-heavy).
Destructive-of-quota: requires confirm='yes' and
ROBOFLOW_MCP_MODE=curate or full. After this returns,
poll roboflow_get_version_generation_status until the
new version reports status="ready".
Inputs: project (required), confirm, dry_run, workspace, +3 more. Returns: object
roboflow_delete_versionA
Delete a dataset version. Destructive: confirm='yes' required.
Inputs: project (required), version (required), confirm, dry_run, +1 more. Returns: result
roboflow_export_versionA
Request an export for a trained/generated version.
Returns a signed download URL when ready. Does NOT stream
bytes — pair with roboflow_download_export if you want
the zip on disk.
Inputs: project (required), version (required), export_format (required), dry_run, +1 more. Returns: result
roboflow_list_annotation_jobsA
List a project's annotation jobs (Annotate-tab batches).
Each job reports labeling / review progress; ready_to_add
marks jobs sitting in review with every image approved, i.e.
eligible for roboflow_add_reviewed_to_dataset. Optionally
filter by status (e.g. "assigned", "review", "complete").
Inputs: project (required), status, dry_run, workspace. Returns: result
roboflow_list_image_batchesC
List upload batches for a project.
Inputs: project (required), dry_run, workspace. Returns: result
roboflow_remove_image_tagsA
Remove tags from an image.
Destructive: requires confirm='yes' and a server mode of
curate or full (see :envvar:ROBOFLOW_MCP_MODE).
Inputs: tags (required), project (required), image_id (required), confirm, +2 more. Returns: object
roboflow_upload_annotationC
Attach an annotation to an image.
Supported formats (annotation_format literal):
"coco""yolo"(YOLOv5/v8/v11-compatible .txt)"pascal_voc"(XML)"createml"(Apple CreateML JSON)"roboflow_json"
Pass annotation as a string (raw file contents) or a dict
(Python object; we serialize to JSON). Optional labelmap
is a class-id to name map Roboflow uses for YOLO formats.
Inputs: project (required), image_id (required), annotation (required), annotation_format (required), +3 more. Returns: result
roboflow_upload_imageA
Upload a single image to a Roboflow project.
source is a discriminated union (exactly one mode):
{"kind": "url", "url": "https://..."}{"kind": "path", "path": "/abs/path/img.jpg"}(local path; must live under ROBOFLOW_MCP_UPLOAD_ROOTS){"kind": "base64", "data": "...", "filename": "img.jpg"}
Every mode runs through URL / path / image safety guards
before the upload. split is "train" / "valid" / "test".
Inputs: source (required), project (required), split, dry_run, +3 more. Returns: result
roboflow_upload_images_batchA
Upload many images concurrently.
sources is a list of :class:ImageSource dicts. Failed
uploads don't abort the batch; the response reports
per-image outcomes in results and errors.
Inputs: project (required), sources (required), split, dry_run, +3 more. Returns: result
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 11 tools
Each tool targets a distinct action or resource: image tagging (add/remove), upload (single/batch), annotation, version management (create/delete/export), listing jobs/batches, and dataset addition. Even similar operations like single vs batch upload are clearly differentiated.
All tools follow a consistent 'roboflow_verb_noun' pattern in snake_case (e.g., roboflow_add_image_tags, roboflow_upload_image). The naming is predictable and easy to parse.
With 11 tools, the set covers the primary operations of a Roboflow MCP server without being bloated. Each tool serves a clear purpose, and the count feels appropriate for the domain.
The tools cover image upload, annotation, tagging, version CRUD (except list/get versions), and dataset addition. However, missing operations include retrieving version details, listing versions, deleting images, and polling version generation status (despite being referenced in create_version). These gaps may hinder full workflow automation.
Maintenance
Related MCP Connectors
Shared memory MCP server for Claude Code, Codex, Gemini CLI and Cursor. Cited, editable.
A comprehensive Model Context Protocol (MCP) server that enables AI assistants to interact with yo…
The OpenRouter MCP server plugs OpenRouter into the AI tools you already use. Once connected, your assistant can pull live OpenRouter data (models, prices, your credits, rankings, and docs) and send quick test messages, all without leaving your editor.
Talk to your public-facing AI from any MCP client — Claude, ChatGPT, Cursor, Cline, Windsurf.
Related MCP Servers
- FlicenseNot gradedqualityDmaintenanceLocal MCP server that provides semantic search (RAG) over code repositories, enabling AI clients like Claude and Gemini to access project context without manual re-upload.-
- AlicenseNot gradedqualityCmaintenanceAn MCP server that exposes your local Claude Code CLI over HTTP+SSE, enabling any MCP-compatible client to use your Claude Code MAX/PRO subscription remotely.24 npm2MIT
- AlicenseNot gradedqualityDmaintenanceA dead simple MCP server for exposing your app functions to AI agents like Claude Desktop.6MIT

Robonine MCP Serverofficial
AlicenseAqualityDmaintenanceA local MCP server that connects Claude Code or any MCP-compatible AI assistant to a Robonine robot arm.1MIT