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., "@mlops-mcp-server track a new experiment with lr=0.01 and log accuracy"
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 for common MLOps workflows
An MCP (Model Context Protocol) server that gives AI assistants like Claude direct access to MLOps workflows — experiment tracking, model registry, dataset management, pipeline orchestration, lineage tracing, and more. Wraps DVC, MLflow, and Git rather than replacing them.
Related MCP server: MLflow MCP Server
Experiment tracking — create runs, log params/metrics/artifacts, compare runs, find best run, export to CSV
Model registry — register, promote, tag, deprecate, and compare model versions; generate model cards
Dataset management — profile datasets, validate schemas, detect statistical drift (KS-test), split, merge, and generate dataset cards
Pipeline management — create, validate, and visualize DAG pipelines as Mermaid diagrams; cycle detection via Kahn's algorithm
Data lineage — record artifact lineage, trace provenance with BFS, visualize as Mermaid graphs, check integrity
File operations — full file/directory CRUD, content search, disk usage, batch operations
Project scaffolding — 7 ML project templates with component injection and structure validation
Documentation generation — model cards, dataset cards, experiment reports, pipeline docs, project READMEs, API docs
MLflow integration — optional; wraps tracking, registry, and artifact operations
DVC integration — optional; wraps data versioning and pipeline reproduction
Git operations — optional; status, add, commit, log, .gitignore generation
Environment tools — scan imports, generate requirements, check conflicts, create conda env files
Tools are registered in two tiers: a small set of always-on tools for quick file ops and discovery, and 15 domain modules you activate on demand to avoid flooding the agent context window.
pip install git+https://github.com/anant-patankar/mlops-mcp-server.gitUsing uv:
uv add git+https://github.com/anant-patankar/mlops-mcp-server.git# Core install
pip install mlops-mcp-server
# With optional extras
pip install mlops-mcp-server[mlflow]
pip install mlops-mcp-server[mlflow,compare,notebooks,parquet]Using uv:
uv add mlops-mcp-server
uv add "mlops-mcp-server[mlflow,compare,notebooks,parquet]"Add to your Claude Desktop MCP config (~/Library/Application Support/Claude/claude_desktop_config.json on macOS):
First install the package, then add to your Claude Desktop config:
{
"mcpServers": {
"mlops": {
"command": "mlops-mcp-server"
}
}
}Restart Claude Desktop. The server starts automatically when Claude connects.
Once connected, you can ask Claude things like:
"Show me all experiment runs, find the one with the best validation accuracy, and register that model in the registry."
"Profile
data/train.csv, check it againstschema.yaml, then split it 70/15/15 and save todata/splits/."
"Create a pipeline with three stages: preprocess → train → evaluate, validate it for cycles, and show me the Mermaid diagram."
mlops-mcp-serverExtra | Installs | Enables |
| mlflow | MLflow tracking and registry tools |
| deepdiff | Structured diff in |
| nbformat | Notebook summary in |
| pyarrow | Parquet read/write in dataset tools |
| joblib, onnx, safetensors | Model file format support |
| pandera, scipy | Schema validation and drift statistics |
| jinja2 | Jinja2 templates in |
All extras are optional — the server runs without them and returns clear errors if a missing dependency is needed.
MIT — see LICENSE.
Inputs: stage_name (required), pipeline_path (required), stage_definition (required). Returns: object
Inputs: component (required), project_path (required). Returns: object
Inputs: path (required). Returns: object
Inputs: experiment_path (required), output_directory, compression_level. Returns: object
Inputs: paths (required). Returns: object
Inputs: paths (required), dry_run, base_path. Returns: object
Inputs: path (required), stale_after_hours. Returns: object
Inputs: project_path. Returns: object
Inputs: template_path (required), env_file_path. Returns: object
Inputs: project_path (required). Returns: object
Inputs: path (required). Returns: object
Inputs: logs_path (required), dry_run, min_size_bytes. Returns: object
Inputs: runs_path (required), dry_run. Returns: object
Inputs: checkpoints_path (required), keep, dry_run. Returns: object
Inputs: project_path (required), dry_run. Returns: object
Inputs: path_a (required), path_b (required). Returns: object
Inputs: path_a (required), path_b (required). Returns: object
Inputs: version_a (required), version_b (required), model_name (required), project_path (required). Returns: object
Inputs: base_path (required), target_path (required). Returns: object
Inputs: run_ids (required), project_path (required). Returns: object
Inputs: source_path (required), destination_path (required), dry_run, base_path. Returns: object
Inputs: source_path (required), archive_path (required), archive_format, compression_level. Returns: object
Inputs: output_path (required), requirements_path (required), env_name, python_version. Returns: object
Inputs: path (required), base_path. Returns: object
Inputs: stages (required), repo_path (required), output_file. Returns: object
Inputs: variables (required), output_path (required). Returns: object
Inputs: path (required), content, encoding, base_path, +1 more. Returns: object
Inputs: overwrite, repo_path. Returns: object
Inputs: project_path (required), template, project_name. Returns: object
Inputs: version (required), model_name (required), project_path (required), output_path. Returns: object
Inputs: stages (required), pipeline_path (required). Returns: object
Inputs: project_path. Returns: object
Inputs: path (required), dry_run, base_path. Returns: object
Inputs: version (required), model_name (required), project_path (required), delete_file. Returns: object
Inputs: run_id (required), project_path (required), dry_run. Returns: object
Inputs: reason (required), version (required), model_name (required), project_path (required). Returns: object
Inputs: candidate_path (required), reference_path (required). Returns: object
Inputs: model_path (required). Returns: object
Inputs: repo_path. Returns: object
Inputs: run_id (required), artifact_path (required), destination_path (required). Returns: object
Inputs: path (required), repo_path. Returns: object
Inputs: no named parameters. Returns: object
Inputs: repo_path. Returns: object
Inputs: repo_path. Returns: object
Inputs: repo_path. Returns: object
Inputs: repo_path. Returns: object
Inputs: repo_path. Returns: object
Inputs: project_path. Returns: object
Inputs: output_path (required), project_path (required). Returns: object
Inputs: archive_path (required), destination_path (required). Returns: object
Inputs: path (required), recursive. Returns: object
Inputs: path (required), encoding. Returns: object
Inputs: path (required), pattern. Returns: object
Inputs: path (required), content (required), encoding. Returns: object
Inputs: path (required). Returns: object
Inputs: dir_path (required), max_size, min_size. Returns: object
Inputs: run_id (required), project_path (required), status. Returns: object
Inputs: source_path (required), output_path. Returns: object
Inputs: dataset_path (required), description, output_path, intended_use, +1 more. Returns: object
Inputs: project_path (required), output_path. Returns: object
Inputs: dataset (required), metrics (required), task_type (required), model_name (required), +3 more. Returns: object
Inputs: pipeline_path (required), output_path. Returns: object
Inputs: project_path (required), output_path. Returns: object
Inputs: output_path, pin_versions, project_path. Returns: object
Inputs: artifact (required), project_path (required). Returns: object
Inputs: metric_name (required), project_path (required), direction. Returns: object
Inputs: data_path (required). Returns: object
Inputs: path, base_path. Returns: object
Inputs: path (required). Returns: object
Inputs: model_name (required). Returns: object
Inputs: experiment_id (required), max_results. Returns: object
Inputs: version (required), model_name (required), project_path (required). Returns: object
Inputs: version (required), model_name (required), project_path (required). Returns: object
Inputs: model_name (required), project_path (required). Returns: object
Inputs: path (required). Returns: object
Inputs: limit. Returns: object
Inputs: pipeline_path (required). Returns: object
Inputs: pipeline_path (required), project_root. Returns: object
Inputs: project_path. Returns: object
Inputs: run_id (required), project_path (required). Returns: object
Inputs: files, repo_path. Returns: object
Inputs: message, repo_path. Returns: object
Inputs: repo_path, create_ignore. Returns: object
Inputs: limit, repo_path. Returns: object
Inputs: repo_path. Returns: object
Inputs: project_path. Returns: object
Inputs: project_path. Returns: object
Inputs: archive_path (required). Returns: object
Inputs: path (required), pattern, base_path, recursive. Returns: object
Inputs: project_path (required). Returns: object
Inputs: no named parameters. Returns: object
Inputs: project_path. Returns: object
Inputs: project_path. Returns: object
Inputs: no named parameters. Returns: object
Inputs: project_path. Returns: object
Inputs: run_id (required), project_path (required), artifact_path (required). Returns: object
Inputs: local_path (required), run_id. Returns: object
Inputs: step (required), run_id (required), metrics (required), project_path (required). Returns: object
Inputs: params (required), run_id (required), project_path (required). Returns: object
Inputs: output_path (required), dataset_paths (required), deduplicate. Returns: object
Inputs: project_path. Returns: object
Inputs: module (required). Returns: object
Inputs: no named parameters. Returns: object
Inputs: no named parameters. Returns: object
Inputs: project_path. Returns: object
Inputs: source_path (required), destination_path (required), dry_run, base_path. Returns: object
Inputs: path (required). Returns: object
Inputs: version (required), new_stage (required), model_name (required), project_path (required). Returns: object
Inputs: path (required), encoding, base_path. Returns: object
Inputs: inputs (required), output (required), run_id (required), project_path (required). Returns: object
Inputs: model_name (required), model_path (required), project_path (required), tags, +3 more. Returns: object
Inputs: model_name (required), model_path (required). Returns: object
Inputs: stage_name (required), pipeline_path (required). Returns: object
Inputs: path (required), new_name (required), dry_run, base_path. Returns: object
Inputs: project_path. Returns: object
Inputs: path (required), query (required), is_regex, base_path, +3 more. Returns: object
Inputs: path (required), max_size, min_size, base_path, +2 more. Returns: object
Inputs: tracking_uri (required), project_path. Returns: object
Inputs: output_dir (required), dataset_path (required), val_ratio, test_ratio, +3 more. Returns: object
Inputs: dir_path (required), top_k.
Inputs: tags (required), version (required), model_name (required), project_path (required). Returns: object
Inputs: schema_path (required), dataset_path (required). Returns: object
Inputs: pipeline_path (required), project_root. Returns: object
Inputs: template (required), project_path (required). Returns: object
Inputs: artifact (required), project_path (required). Returns: object
Inputs: pipeline_path (required). Returns: object
Inputs: path (required), content (required), encoding, base_path. Returns: object
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 127 tools
With 127 tools and no descriptions, many tool names overlap significantly (e.g., file_read vs read_file, list_models vs model_list, etc.). An agent would be unable to distinguish between them, leading to frequent misselection.
Naming is highly inconsistent, mixing verb_noun, noun_verb, and prefixes like mlops_, mlflow_, dvc_, git_. Similar operations use different naming patterns (e.g., file_read vs read_file), making it unpredictable.
127 tools is far too many for a coherent server. Typical MCP servers have 3-15 tools. This large number suggests an attempt to cover everything, resulting in bloat and redundancy rather than a focused tool set.
Despite the large number of tools, the surface appears redundant (multiple file operations, repeated model list functions) and lacks clear high-level operations like deploy or monitor. Without descriptions, it's hard to assess, but there are likely gaps in ML Ops lifecycle.
Query your team's drift, vulnerability, and upgrade data from any AI assistant. OAuth 2.1, 51 tools.
Git-backed platform for skills, tools, and context for AI agents
Connect AI assistants to GitHub - manage repos, issues, PRs, and workflows through natural language.
Plain-English git for agents: 22 tools to branch, commit, push and tag. Free tier.
