Clearml Codex Plugin

ClearML Plugin

Read-only ClearML MCP tools for Codex. The plugin uses the official ClearML Python SDK backend API and never modifies tasks, artifacts, models, queues, or projects.

Credentials are read from the normal ClearML SDK locations: CLEARML_CONFIG_FILE, ~/clearml.conf, ~/.clearml.conf, or the standard ClearML environment variables. Do not pass secrets in prompts or tool arguments.

Tools

• clearml_search_tasks - find tasks by project id/name, task name filter, tags, status, and limit.
• clearml_project_snapshot - summarize a project/search slice with task metadata and optional latest metrics.
• clearml_experiment_slice - export a compact slice for selected experiments: metadata, discovered metrics, latest scalars, optional equal-step compare, series tails, logs, and raw events.
• clearml_metrics - list scalar metric names and variants for tasks.
• clearml_latest_scalars - fetch latest scalar values for tasks.
• clearml_scalar_series - fetch one scalar metric/variant series.
• clearml_compare_tasks - compare tasks at a shared step using explicit metrics or a preset.
• clearml_compare_url - build a ClearML web comparison URL.
• clearml_task_info - fetch task metadata.
• clearml_task_logs - fetch recent task logs.
• clearml_task_events - fetch raw task events.

Skill

The plugin includes skills/clearml/SKILL.md. It tells Codex to:

• search for tasks first when ids are missing;
• compare runs at the same step by default;
• lead domain-specific reports with outcome metrics, not only loss curves;
• use Browser tools to open ClearML UI pages when they are available;
• otherwise return ClearML URLs directly.

Metric Paths

Scalar paths use ClearML metric/variant form, for example train/loss or sampling_val/fcd.

Presets

Presets are conveniences only; the API is generic.

• training-basic
• generation-basic

Example Requests

Find runs by tag in a project:

Search ClearML tasks in project  with tag .

Compare known runs at equal progress:

Compare these ClearML tasks at the same step using an explicit metric_specs map or a preset.

Export a compact experiment slice:

Get an experiment slice for these task ids with latest scalars, sampling_val/fcd series tail, logs tail, and a compare URL.

Open or return a comparison page:

Build the ClearML compare page for these task ids. If Browser tools are available, open it and inspect the charts.

Local Test

python -m pytest tests -q
uv run --with clearml --with mcp python -c "import mcp_server"

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: SergeiNikolenko
• Source: SergeiNikolenko/clearml-codex-plugin
• License: MIT

Install and usage instructions live in the source repository linked above.