CVAT MCP
This server enables natural language interaction with a CVAT (Computer Vision Annotation Tool) instance by exposing its REST API as MCP tools usable by AI assistants like Cline.
Server Info: Retrieve basic CVAT server information and inspect the full OpenAPI schema.
Task Management: List tasks with filters (search, status, project, assignee, pagination); get a specific task by ID; create new tasks with labels and configuration; attach media (local files, remote URLs, or server-side files).
Annotation Management: Retrieve annotations for a task; replace/overwrite task annotations (requires explicit confirmation).
Job Management: List jobs with filters (task, project, state, stage, assignee, pagination); get a specific job by ID.
Project Management: List projects with filters; create new projects with labels.
Label Management: List labels filtered by task, project, job, or name.
Background Request Tracking: Poll the status of async CVAT operations (imports, exports, uploads).
Generic API Access: Call any CVAT REST endpoint directly (GET, POST, PATCH, PUT, DELETE), with mutation operations requiring explicit confirmation.
Extended Tool Generation: Generate additional MCP tools from the CVAT OpenAPI schema to cover version-specific or less common endpoints.
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., "@CVAT MCP List tasks with status 'annotation' and their assignees"
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.
CVAT MCP
CVAT MCP는 Cline에서 자연어로 CVAT 작업을 수행할 수 있게 하는 stdio MCP 서버입니다. CVAT의 공식 REST API만 호출하며, Node 내장 기능만 사용하므로 별도 런타임 의존성이 없습니다.
목표
CVAT 공식 API를 MCP Tool로 노출한다.
Cline에서 "CVAT task 목록 보여줘", "새 task 만들고 이미지를 업로드해" 같은 텍스트 요청을 하면 적절한 CVAT API가 호출되도록 한다.
기존 CVAT 프로젝트에 어떻게 포함될지에 대한 단계별 계획을 문서로 남긴다.
Related MCP server: WorkItems DevOps MCP Server
구성
src/index.js: Cline이 실행하는 MCP stdio 엔트리포인트src/static-tools.js: 자주 쓰는 CVAT API를 고수준 MCP Tool로 정의src/cvat-client.js: CVAT REST API 클라이언트scripts/generate-tool-catalog.js: CVAT OpenAPI 스키마에서 추가 MCP Tool 카탈로그 생성docs/integration-plan.md: 기존 CVAT에 포함시키는 계획docs/cvat-api-tooling.md: CVAT API를 Tool로 확장하는 방식
요구사항
Node.js 20 이상
접근 가능한 CVAT 서버
CVAT Personal Access Token 권장
CVAT 공식 문서는 서버 API가 HTTP REST API이며, API 스키마와 Swagger 문서를 서버에서 제공한다고 설명합니다. PAT 인증은 Authorization: Bearer <token> 헤더를 사용합니다.
실행
npm test$env:CVAT_BASE_URL="http://localhost:8080"
$env:CVAT_AUTH_TOKEN="your_pat"
npm start서버는 MCP stdio 프로세스이므로 사람이 직접 대화형으로 쓰는 프로그램이 아닙니다. Cline이 프로세스를 실행하고 stdin/stdout으로 JSON-RPC 메시지를 주고받습니다.
Cline 설정
Cline의 cline_mcp_settings.json에 아래 서버를 추가합니다. Windows 경로는 실제 저장소 경로로 바꿉니다.
{
"mcpServers": {
"cvat-mcp": {
"command": "node",
"args": ["C:\\Users\\준승\\Desktop\\Development\\CVAT-MCP\\src\\index.js"],
"env": {
"CVAT_BASE_URL": "http://localhost:8080",
"CVAT_AUTH_TOKEN": "your_pat",
"CVAT_AUTH_SCHEME": "Bearer"
},
"disabled": false,
"alwaysAllow": [
"cvat_server_about",
"cvat_list_tasks",
"cvat_get_task",
"cvat_list_projects",
"cvat_list_jobs"
]
}
}
}쓰기 작업을 자동 승인하려면 신중해야 합니다. cvat_api_request의 POST, PATCH, PUT, DELETE와 생성된 mutation 도구는 confirmMutation: true가 있어야 실행됩니다.
Cline 프롬프트 예시
CVAT 서버 상태를 확인해줘.CVAT에서 "road"가 들어간 task를 찾아서 상태와 assignee를 요약해줘.CVAT에 "sample cars" task를 만들고 label은 car, person으로 설정해줘.task 12의 annotation JSON을 가져와서 현재 workspace에 요약해줘.기본 제공 Tool
cvat_server_aboutcvat_get_api_schemacvat_list_taskscvat_get_taskcvat_create_taskcvat_attach_task_datacvat_get_task_annotationscvat_replace_task_annotationscvat_list_jobscvat_get_jobcvat_get_requestcvat_list_projectscvat_create_projectcvat_list_labelscvat_api_request
공식 API 전체를 Tool로 확장
CVAT 서버에서 OpenAPI 스키마를 받아 추가 도구 카탈로그를 만들 수 있습니다.
node scripts/generate-tool-catalog.js http://localhost:8080/api/schema/ generated/cvat-tools.json인터넷 연결은 필수가 아닙니다. MCP 서버와 생성 스크립트는 CVAT_BASE_URL 또는 명령에 입력한 CVAT 주소로 직접 API를 호출합니다. 내부망 CVAT 서버를 쓰는 경우에도 MCP 서버가 그 서버의 /api/schema/에 접근할 수 있으면 Tool 카탈로그를 생성할 수 있습니다.
다만 CVAT 버전에 따라 OpenAPI 스키마와 엔드포인트가 다를 수 있습니다. https://app.cvat.ai/api/schema/처럼 외부 공개 서버 기준으로 생성한 Tool은 내부 CVAT 서버 버전과 맞지 않을 수 있으므로, 실사용 전에는 실제 사용할 CVAT 서버의 schema로 다시 생성하는 것을 권장합니다.
node scripts/generate-tool-catalog.js http://실제_CVAT_주소/api/schema/ generated/cvat-tools.json이후 Cline 설정에 아래 환경 변수를 추가합니다.
{
"CVAT_GENERATED_TOOLS_PATH": "generated/cvat-tools.json"
}생성 도구는 {path, query, body} 형태의 입력을 받고 실제 CVAT 엔드포인트로 전달됩니다. 자주 쓰는 작업은 정적 고수준 도구를 우선 사용하고, CVAT 버전별 새 엔드포인트는 생성 도구 또는 cvat_api_request로 처리합니다.
참고 문서
CVAT Server API: https://docs.cvat.ai/docs/api_sdk/api/
CVAT Access Tokens: https://docs.cvat.ai/docs/api_sdk/access_tokens/
Cline MCP 설정: https://docs.cline.bot/mcp/adding-and-configuring-servers
MCP stdio transport: https://modelcontextprotocol.io/specification/2025-06-18/basic/transports
Available Tools
15 toolscvat_api_requestB
Call any official CVAT REST endpoint under /api/. For POST, PATCH, PUT, or DELETE, confirmMutation must be true.
Inputs: method (required), path (required), query, body, +2 more.
cvat_attach_task_dataB
Attach local, server, or remote media to a task using POST /api/tasks/{id}/data/ with CVAT upload headers.
Inputs: taskId (required), clientFiles, serverFiles, remoteFiles, +1 more.
cvat_create_projectC
Create a CVAT project using POST /api/projects.
Inputs: name (required), labels (required), bug_tracker.
cvat_create_taskB
Create a CVAT task using POST /api/tasks. Use cvat_attach_task_data afterward to upload images or videos.
Inputs: name (required), labels (required), project_id, assignee_id, +3 more.
cvat_get_api_schemaA
Fetch the CVAT OpenAPI schema from GET /api/schema/ so official API endpoints can be inspected or converted into MCP tools.
Inputs: no named parameters.
cvat_get_jobB
Get one CVAT job by id from GET /api/jobs/{id}.
Inputs: jobId (required).
cvat_get_requestA
Get a CVAT background request by id from GET /api/requests/{id}; useful after imports, exports, or data upload.
Inputs: requestId (required).
cvat_get_taskC
Get one CVAT task by id from GET /api/tasks/{id}.
Inputs: taskId (required).
cvat_get_task_annotationsD
Get task annotations from GET /api/tasks/{id}/annotations/.
Inputs: taskId (required).
cvat_list_jobsC
List CVAT jobs with official /api/jobs filters such as task_id, project_id, state, stage, assignee, sort, page, and page_size.
Inputs: task_id, project_id, state, stage, +4 more.
cvat_list_labelsC
List labels with /api/labels filters such as task_id, project_id, job_id, name, search, and page_size.
Inputs: task_id, project_id, job_id, name, +3 more.
cvat_list_projectsC
List CVAT projects with /api/projects filters such as search, owner, assignee, sort, page, and page_size.
Inputs: search, owner, assignee, sort, +2 more.
cvat_list_tasksC
List CVAT tasks with official /api/tasks filters such as search, status, project_id, assignee, owner, sort, page, and page_size.
Inputs: search, status, project_id, assignee, +4 more.
cvat_replace_task_annotationsC
Replace task annotations with PUT /api/tasks/{id}/annotations/. Requires confirmReplace=true.
Inputs: taskId (required), annotations (required), confirmReplace (required).
cvat_server_aboutA
Get basic CVAT server information from GET /api/server/about.
Inputs: no named parameters.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 15 tools
Each tool targets a distinct action and resource. The generic 'cvat_api_request' is clearly a fallback for uncovered endpoints, and all specific tools have unique purposes (create, list, get, attach, replace, etc.) without overlap.
All tools follow a consistent 'cvat_verb_noun' pattern (e.g., cvat_create_task, cvat_list_jobs). The convention is uniform across the entire set, with no mixing of styles.
15 tools is well within the ideal 3-15 range. The number is appropriate for a CVAT server covering projects, tasks, jobs, labels, annotations, and server introspection without being overwhelming.
The set covers core creation, listing, and annotation operations but lacks dedicated update and delete tools for projects and tasks. A generic API request tool exists but requires API knowledge, leaving notable gaps for typical agent workflows.
Maintenance
Related MCP Connectors
Manage Redbooth tasks, task lists, comments, tags and files from Claude and other AI assistants.
Manage projects, tasks, time tracking, and team collaboration through natural language.
Connects AI assistants to TrackingTime for querying time entries, projects, tasks, customers, and team data using natural language.
Connect your Cormak.app account to manage projects, task boards, AI agents and creative work from your AI assistant. The server exposes up to 83 tools, limited to the OAuth permissions you choose. Organize tasks, read and update project drafts, inspect saved studio results, create product descriptions and images, and request supported photo, video, 3D, voiceover, SEO and advertising operations. Publishing saved projects requires separate owner approval. Reading and ordinary manual task management do not themselves start paid generation. AI operations use credits from your connected Cormak account and require owner confirmation, or an existing bounded spending permission for supported product operations. You can limit and revoke connections in Cormak. The MCP does not provide bank access, payment details or provider API keys. CRM, streaming website generation and business-plan generation are not exposed through this integration. Get started: https://app.cormak.app/mcp Setup guide and examples: https://cormak.app/blog/cormak-mcp-guide Privacy: https://cormak.app/privacy Support: https://cormak.app/contact
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceEnables AI assistants to interact with OpenProject's API v3 for comprehensive project management operations including work packages, projects, time tracking, users, and all other OpenProject features through natural language.4MIT
- FlicenseNot gradedqualityFmaintenanceEnables LLMs and AI applications to interact with Azure DevOps Work Items, supporting queries, filtering, status updates, date management, effort tracking, descriptions, and comments through natural language.-
- AlicenseBqualityDmaintenanceEnables interaction with Vikunja task management instances through natural language. Supports comprehensive project and task operations including CRUD, assignments, labels, comments, relations, and attachments.3343 npm2MIT
- AlicenseCqualityDmaintenanceEnables AI assistants to interact with Vikunja task management instances, providing full task, project, label, and user management capabilities.17214 npmMIT