Roboflow

Name: Roboflow MCP Server
Availability: InStock
Author: Roboflow

HIPAA CompliantSOC2 ReadyISO 27001 Ready

Developer Tools67 tools

Manage computer vision datasets, train models, run inference and build Workflows. 67 tools and 63 per-resource scopes — the finest-grained consent here.

Verified connector

Listed by Anthropic as a partner connector in its Connectors Directory.

Connection checked by Agentman on 18 August 2026.

Anthropic states this reflects the level of review a connector received, not a security audit.

Connect Roboflow via MCP

https://mcp.roboflow.com/mcp

Works in any MCP-compatible client. In Agentman Studio it is one click — no config file to edit.

Use in Agentman

Connect Roboflow once and your agents call these tools on their own — on a schedule, in a workflow, with nobody at the keyboard.

Open in Agentman Studio

Roboflow Tools & Capabilities (67)

agent_chat

agent_conversation_get

agent_conversations_list

agent_workflow_publish

annotation_batches_get

annotation_batches_list

annotation_jobs_create

annotations_save

async_tasks_get

autolabel_job_get

autolabel_start

devices_create

devices_get

devices_get_config

devices_get_config_history

devices_get_default_config

devices_get_events

devices_get_logs

devices_get_telemetry

devices_list

devices_streams_get

devices_streams_list

devices_update_config

image_upload

image_upload_status

images_batch_update_metadata

images_search

images_update_metadata

images_workspace_search

meta_feedback_send

model_evals_get

model_evals_get_confidence_sweep

model_evals_get_confusion_matrix

model_evals_get_image_predictions

model_evals_get_map_results

model_evals_get_performance_by_class

model_evals_get_recommendations

model_evals_get_vector_analysis

model_evals_list

models_get

models_get_training_status

models_infer

models_list

models_star_nas

models_train

projects_create

projects_fork

projects_get

projects_health

projects_list

trainings_cancel

trainings_get_results

trainings_stop

universe_dataset_images_search

universe_search

versions_export

versions_generate

versions_get

workflow_blocks_get_schema

workflow_blocks_list

workflow_specs_run

workflow_specs_validate

workflows_create

workflows_get

workflows_list

workflows_run

workflows_update

Tool names from Anthropic's directory listing. This server requires sign-in, so we could not read tool descriptions or parameter schemas.

Limits

This connector can spend money. Roboflow charges credits for data, training and deployment, and its documentation says credits are consumed whether the feature runs locally or on Roboflow's servers. models_train, models_infer, workflows_run and autolabel_start all bill. Flex billing is on by default for every workspace with a 100 dollar per-month cap you can raise, lower or remove — set that cap deliberately before connecting an agent.
Set a credit cap on training. Roboflow's training documentation describes a per-run credit cap that stops the job and keeps the weights when reached. The cap requires a plan with usage-based billing, and a value below the model's minimum is rejected with invalid_credit_cap.
The scope catalogue is wider than the tool list. 22 of the 63 application scopes — including the whole api-key: and credentials: families — match no tool in either the directory listing or Roboflow's reference, as of the dates above. Grant narrowly.
The connector cannot permanently delete. Roboflow reserves permanent deletion to the web app, explicitly so a stray script or automation cannot irrecoverably destroy data. Deleted items sit in Trash for 30 days.
Non-Admin roles can do very little. Under Roboflow's role model, Reviewers and Labelers cannot train models, build Workflows, deploy models, upload or delete images, or view API keys. Role-based access control is itself a premium feature.
Some architectures are plan-gated. Roboflow's training documentation restricts RF-DETR Medium, Large and Extra Large to paid users, and SAM3 to paid plans with usage-based billing. A training request for an unavailable architecture returns HTTP 400 and starts no job.
We could not read tool schemas or safety annotations. The endpoint requires authentication, so the 67 count and the names are the directory's, and parameter-level detail is not published here. We never authenticated, called a tool, or started a training or inference job.
Five tool names differ between Roboflow's docs and the directory, in each direction, as of the dates above.
The connector is not the only Roboflow agent. Roboflow also ships an in-app Roboflow Agent and a separate Agent Skills package installed with npx @roboflow/skills install. Setup steps for one do not apply to the others.

Frequently asked questions

No. Roboflow's MCP documentation states the server uses OAuth for authentication and that no API key is needed, prompting you to sign in to Roboflow on first use. We confirmed the posture on 2026-08-18: an anonymous request to the endpoint returned HTTP 401 with an OAuth challenge, so nothing connects without a browser consent step.

Yes. Roboflow bills training, inference and Workflow runs in credits, and the connector exposes models_train, models_infer and workflows_run. Roboflow's credits documentation says flex billing is enabled by default for all workspaces with a 100 dollar monthly cap, so an agent can consume credits past your plan allocation unless you lower or disable that cap.

Any plan, including the free Public Plan, but the useful actions need an Admin role. Roboflow's access-control documentation reserves training models, building Workflows, deploying models, uploading and deleting images, and viewing API keys to Admins, and notes that without the premium role-based access control feature every user is an Admin anyway.

Not through the API it uses. Roboflow's Trash documentation states permanent deletion is available only from the web app and is not exposed on the REST API, Python SDK or CLI, so a stray script cannot irrecoverably destroy data. Deleted projects, versions, trainings and Workflows sit in Trash for 30 days and can be restored.

No. Roboflow's Trash documentation states that in-flight training jobs for a project or version are cancelled automatically when the item is moved to Trash, so you do not keep spending credits on something you are deleting. The connector also exposes trainings_stop and trainings_cancel for stopping a run directly.

It returns HTTP 429 and you should back off. Roboflow's production readiness checklist tells callers to retry 429 and 5xx responses with exponential backoff and jitter, and never to retry 400, 401, 403 or 404 automatically. Device log reads carry an explicit limit of 5 requests per minute per IP.