mcp-clearml 0.1.0

MCP server for ClearML

MCP server for ClearML: browse projects, search and compare tasks, list datasets and models — all via Model Context Protocol.

• PyPI: https://pypi.org/project/mcp_clearml/
• License: MIT
• Docs: https://mcp-clearml.readthedocs.io

---

What you get

This server exposes a curated set of ClearML operations as MCP tools:

• Projects
  • list_of_all_projects: list all projects (id, name)
  • find_project_by_pattern: find by name substring (case-insensitive)
  • get_project_stats: aggregated counters for a project (statuses, users, tags, days)
• Tasks (bulk/filters)
  • get_tasks_core_info: core fields by task ids
  • get_tasks_full_info: core + parameters/metrics/artifacts/models by task ids
  • find_tasks_core_info_by_pattern: IDs by name/status/tags
  • find_tasks_full_info_by_pattern: full profiles by name/status/tags
  • get_tasks_core_info_by_project: stats + core list for a project
  • get_tasks_full_info_by_project: stats + full profiles for a project
• Models
  • find_models_by_pattern: search models by name fragment (includes url/uri when present)
  • find_models_info: info for specific model ids
• Datasets
  • find_datasets_by_project: list datasets in a project (optional recursive)
  • find_datasets_by_pattern: list datasets by partial name
  • get_datasets_full_info: per-dataset sizes, uploader and parsed metadata (csv/csv.gz/json)
• Compare
  • compare_tasks: cross-task summary + aligned metrics + parameters diff

See also in-code guides to help LLMs choose the right tool:

• Categories: mcp_clearml.docs.CATEGORY_GUIDE
• Per-tool hints: mcp_clearml.docs.TOOL_GUIDE

---

Requirements

• Python >= 3.12
• ClearML account with valid credentials in ~/.clearml/clearml.conf
• Recommended uv to run via uvx without installing globally

---

Quick Start

Prerequisites

Ensure your ~/.clearml/clearml.conf contains your credentials:

[api]
api_server = https://api.clear.ml
web_server = https://app.clear.ml
files_server = https://files.clear.ml

credentials {
    "access_key": "your-access-key",
    "secret_key": "your-secret-key"
}

You can obtain keys in ClearML Settings.

Install / Run

• Install from PyPI:

pip install mcp_clearml

• Run without installation (via uvx):

uvx mcp-clearml

• Local dev:

uv sync
uv run mcp-clearml

---

Run MCP server (stdio)

The mcp-clearml command validates ClearML connectivity and starts an MCP stdio server.

---

Integrations (MCP clients)

Claude Desktop

Config file:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "clearml": {
      "command": "uvx",
      "args": ["mcp-clearml"]
    }
  }
}

Alternative (if installed via pip):

{
  "mcpServers": {
    "clearml": {
      "command": "python",
      "args": ["-m", "mcp_clearml.mcp"]
    }
  }
}

Cursor

Settings → MCP → Add Server:

{
  "mcp.servers": {
    "clearml": {
      "command": "uvx",
      "args": ["mcp-clearml"]
    }
  }
}

You can also add a rule in .cursorrules to remind using the clearml MCP server for experiment analysis.

Continue

~/.continue/config.json:

{
  "mcpServers": {
    "clearml": {
      "command": "uvx",
      "args": ["mcp-clearml"]
    }
  }
}

Cody

{
  "cody.experimental.mcp": {
    "servers": {
      "clearml": {
        "command": "uvx",
        "args": ["mcp-clearml"]
      }
    }
  }
}

Any MCP‑compatible assistant

{
  "mcpServers": {
    "clearml": {
      "command": "uvx",
      "args": ["mcp-clearml"]
    }
  }
}

Verified with: Zed, OpenHands, Roo‑Cline, and others.

---

Use in Claude Code (or any MCP client)

Add the stdio server with the same command. Then call tools by name (e.g., "find tasks by pattern", "compare_tasks"). Generic MCP UIs will connect over stdio automatically.

---

Examples

• "List projects" → list_of_all_projects
• "Stats for project X" → get_project_stats { project_name: "X" }
• "Find tasks by name fragment and get full details" →
  1. find_tasks_core_info_by_pattern { task_name_pattern: "resnet" }
  2. get_tasks_full_info { task_ids: ["..."] }
• "Compare two tasks" → compare_tasks { task_ids: ["task_id_1", "task_id_2"] }
• "Find datasets by project" → find_datasets_by_project { project_name: "X", recursive_project_search: true }

---

Development

Setup & test locally:

uv sync --extra test
uv run pytest -q

Coverage gate is configured at 65% (see [tool.coverage.*] in pyproject.toml).

Lint/type check:

uv run ruff check --output-format=github src/ tests/
uv run ty check || true

---

Release (GitHub Actions)

• Tests run on all branches/PRs (.github/workflows/tests.yml).
• Tag‑based release (.github/workflows/release.yml):
  1. push tag vX.Y.Z → runs tests (verify)
  2. build wheels/sdist, create GitHub Release with artifacts
  3. optional PyPI publish if PYPI_API_TOKEN is set

See docs/releasing.md for the exact steps.

---

Credits

Acknowledgment: thanks to prassanna-ravishankar for the ClearML MCP project that inspired this work.