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., "@Rockfish MCP Server start training a TabGAN model on the 'sales' dataset"
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.
A Model Context Protocol server that provides tools to interact with the Rockfish AI platform for synthetic data generation, dataset management, and ML workflow orchestration.
Rockfish API — databases, worker sets, workflows, models, projects, datasets, organizations (22+ tools)
SDK (Synthetic Data Generation)
obtain_train_config — generate training configuration with automatic column type detection
update_train_config — modify training hyperparameters or field classifications
start_training_workflow — start TabGAN training workflow
get_workflow_logs — stream workflow logs with configurable level and timeout
get_trained_model_id — extract trained model ID from completed workflow
start_generation_workflow — start generation workflow from trained model
obtain_synthetic_dataset_id — extract generated dataset ID from completed workflow
plot_distribution — generate distribution plots comparing real and synthetic data
get_marginal_distribution_score — calculate similarity score between real and synthetic data
Manta (Analytics & Scenarios) — requires MANTA_API_URL
discover_schema — discover dataset schema
generate_test_suite — generate test suites
execute_query / execute_nl_query — run SQL or natural language queries
inject_scenario — inject test scenarios into datasets
rockfish-mcp is not currently published to PyPI, so install it from source (below) and run it
with uv run --directory /path/to/rockfish-mcp rockfish-mcp. The uvx rockfish-mcp and
pip install rockfish-mcp forms only work once the package is published to an index.
git clone https://github.com/Rockfish-Data/rockfish-mcp.git
cd rockfish-mcpInstall dependencies (exact versions from uv.lock, including dev tools):
uv sync
source .venv/bin/activate # optional; or prefix commands with `uv run`For a runtime-only install (no dev tools), use uv sync --no-dev.
Set up environment variables:
cp .env.example .env
# Edit .env and add your Rockfish API keyRelated MCP server: RAGFlow MCP Server
Create a .env file with your Rockfish API credentials:
ROCKFISH_API_KEY=your_api_key_here
ROCKFISH_API_URL=https://api.rockfish.aiOptional settings:
ROCKFISH_ORGANIZATION_ID=your_organization_id_here
ROCKFISH_PROJECT_ID=your_project_id_here
MANTA_API_URL=https://manta.rockfish.aiAdd to your claude_desktop_config.json:
rockfish-mcp is not published to PyPI, so uvx rockfish-mcp fails with
rockfish-mcp was not found in the package registry. Run it from a clone instead. Use
uv run --directory rather than uvx --from, so the locked dependencies and the Rockfish SDK
find-links index from pyproject.toml are honored. Use the absolute path to uv
(which uv), since Claude Desktop does not inherit your shell PATH.
{
"mcpServers": {
"rockfish": {
"command": "/opt/homebrew/bin/uv",
"args": ["run", "--directory", "/path/to/rockfish-mcp", "rockfish-mcp"],
"env": {
"ROCKFISH_API_KEY": "your_api_key_here",
"ROCKFISH_API_URL": "https://api.rockfish.ai"
}
}
}
}The first launch builds the environment and can take ~30 seconds; later launches are fast.
{
"mcpServers": {
"rockfish": {
"command": "uvx",
"args": ["rockfish-mcp"],
"env": {
"ROCKFISH_API_KEY": "your_api_key_here",
"ROCKFISH_API_URL": "https://api.rockfish.ai"
}
}
}
}{
"mcpServers": {
"rockfish": {
"command": "python",
"args": ["-m", "rockfish_mcp.server"],
"env": {
"ROCKFISH_API_KEY": "your_api_key_here",
"ROCKFISH_API_URL": "https://api.rockfish.ai"
}
}
}
}{
"mcpServers": {
"rockfish": {
"command": "docker",
"args": ["run", "-i", "--rm", "-e", "ROCKFISH_API_KEY", "-e", "ROCKFISH_API_URL", "rockfish-mcp"],
"env": {
"ROCKFISH_API_KEY": "your_api_key_here",
"ROCKFISH_API_URL": "https://api.rockfish.ai"
}
}
}
}You can use the MCP inspector to debug the server from a source checkout:
npx @modelcontextprotocol/inspector uv run --directory path/to/rockfish-mcp rockfish-mcpOr run the venv's Python directly (after uv sync):
cd path/to/rockfish-mcp
npx @modelcontextprotocol/inspector .venv/bin/python -m rockfish_mcp.serverIf the server fails to start in Claude Desktop, check ~/Library/Logs/Claude/mcp-server-rockfish.log.
Install with dev dependencies (the default for uv sync):
uv syncRun tools with uv run (for example uv run pytest) or activate .venv first.
isort src/rockfish_mcp/ && black src/rockfish_mcp/Unit tests (no credentials required):
pytest tests/test_manta_client.py tests/test_manta_tools.pyIntegration tests (requires .env with real credentials):
pytest tests/
pytest --env=.env.prod.local --html=report.html --self-contained-htmlWe encourage contributions to help expand and improve rockfish-mcp. Whether you want to add new tools, enhance existing functionality, or improve documentation, your input is valuable.
Fork the repository
Create a feature branch
Make your changes
Format your code with isort and black
Add tests if applicable
Submit a pull request
For examples of other MCP servers and implementation patterns, see: https://github.com/modelcontextprotocol/servers
rockfish-mcp is licensed under the MIT License. This means you are free to use, modify, and distribute the software, subject to the terms and conditions of the MIT License. For more details, please see the LICENSE file in the project repository.
Create a new database
Inputs: name (required), description.
Create a new dataset
Inputs: name (required), description.
Create a new project
Inputs: name (required), description.
Create a new worker set
Inputs: name (required), worker_count (required).
Create and run a new workflow
Inputs: jobs (required), metadata (required).
Delete a database
Inputs: id (required).
Delete a dataset
Inputs: id (required).
Delete a model
Inputs: id (required).
Delete a worker set
Inputs: id (required).
Execute a SQL query against Rockfish datasets and return results in CSV format
Inputs: query (required), project_id.
Get active organization
Inputs: no named parameters.
Get active project
Inputs: no named parameters.
Get a specific database by ID
Inputs: id (required).
Get a specific dataset by ID
Inputs: id (required).
Retrieve the query bookmarks for a dataset
Inputs: dataset_id (required).
Get the schema for a dataset
Inputs: id (required).
Calculate marginal distribution score between two datasets (real vs synthetic). Lower score means more similar distributions.
Inputs: dataset_ids (required).
Get a specific model by ID
Inputs: id (required).
Get a specific project by ID
Inputs: id (required).
Extract the trained model ID from a completed training workflow. Only works on COMPLETED or FINALIZED workflows.
Inputs: workflow_id (required).
Get a specific worker set by ID
Inputs: id (required).
List the actions for a worker-set
Inputs: id (required).
Get a specific workflow by ID
Inputs: id (required).
Stream logs from a workflow execution with configurable log level and collection timeout. Collects logs for specified duration (default 10s), useful for monitoring workflow progress and debugging. Call repeatedly to continue collecting logs from running workflows.
Inputs: workflow_id (required), timeout, log_level.
List actions (or workers) from all worker-sets
Inputs: no named parameters.
List all databases
Inputs: no named parameters.
List all datasets
Inputs: no named parameters.
List all models
Inputs: no named parameters.
List all organizations
Inputs: no named parameters.
List all projects
Inputs: no named parameters.
List all worker groups
Inputs: no named parameters.
List all worker sets
Inputs: no named parameters.
List all workflows
Inputs: no named parameters.
Extract the synthetic dataset ID from a completed generation workflow. Only works on COMPLETED or FINALIZED workflows.
Inputs: generation_workflow_id (required).
Generate training configuration for synthetic data models. Returns a cached train_config_id for use in start_training_workflow. Supports TabGAN (rf_tab_gan) and TimeGAN (rf_time_gan).
Inputs: dataset_id (required), model_type (required).
Generate distribution plot comparing datasets for a specific column. Returns base64-encoded PNG image. Automatically chooses bar plot (categorical) or KDE plot (numerical).
Inputs: column_name (required), dataset_ids (required).
Start a workflow to generate synthetic data from a trained TabGAN model.
Inputs: model_id (required).
Build and start a training workflow using a cached config. Automatically detects model type from train_config structure. Use this after obtaining train_config_id from obtain_train_config.
Inputs: dataset_id (required), train_config_id (required).
Update a database
Inputs: id (required), name, description.
Update a dataset
Inputs: id (required), name, description.
Update a project
Inputs: id (required), name, description.
Update training configuration. Modify model hyperparameters (epochs, batch_size, etc.) and/or encoder field classifications (categorical, continuous, ignore). See https://docs.rockfish.ai/model-train.html for config structure.
Inputs: updates (required), train_config_id (required).
Update a workflow
Inputs: id (required), name, definition.
Upload a new model
Inputs: name (required), metadata.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 44 tools
Most tools target a distinct resource and action (e.g., create_dataset, get_workflow, start_training_workflow). A few listing tools overlap in scope, such as list_available_actions versus get_worker_set_actions, and list_worker_groups versus list_worker_sets, but descriptions generally clarify the differences.
Nearly all tools follow a consistent snake_case verb_noun pattern. Minor deviations include using obtain instead of get for extraction tools and upload_model instead of create_model, but the overall convention remains predictable.
With 44 tools, the surface is heavy for the apparent synthetic-data platform scope. While the domain is broad, many operations could likely be consolidated or grouped, and 25+ tools is generally too many for an agent to navigate efficiently.
Core synthetic data lifecycle is well covered: dataset CRUD, SQL query, model upload, training config, training workflow, generation workflow, and evaluation plots. Minor gaps exist, such as missing delete_project, delete_workflow, and update_model, but agents can work around these.
Run any Relay model (image, video, audio, LLM), poll tasks, upload files and search Relay's docs.
List datasets, schemas, run APL queries, and use prompts for exploration, anomalies, and monitoring.
AI-powered data integration platform. Onboard users and run DPF data workflows.
Access Oi Contexts, Workflows, Skills, Guardrails, Connections, and reporting tools.
