The config file is at:

- macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

- Windows: %APPDATA%\Claude\claude_desktop_config.json

Then restart Claude Desktop.
Claude Code

Other MCP clients

Environment variables

| Variable | Required | Description |

|---|---|---|

| PROXYML_API_KEY | Yes | Your ProxyML API key |

Security & permissions
What this server can access
Filesystem — proxyml_infer_schema reads one CSV file at the path you

provide. No other tool reads or writes local files.

Network — outbound HTTPS only, to https://api.proxyml.ai (or the URL

set in PROXYML_BASE_URL). No other hosts are contacted.

Environment — only PROXYML_API_KEY and PROXYML_BASE_URL are read.

No other environment variables are accessed.

No shell execution, no local ports, no other system resources.
What is sent to ProxyML's servers
| Tool | What is transmitted |

|---|---|

| proxyml_infer_schema | Nothing — CSV is read and processed locally only |

| proxyml_put_schema | Feature schema: column names, types, and constraints |

| proxyml_get_schema | Schema name (string) |

| proxyml_synthesize_data | Schema name, sample count, and optionally one feature vector |

| proxyml_train_surrogate | Synthetic feature vectors + your model's predictions for those vectors |

| proxyml_predict_batch | Feature vectors you pass explicitly |

| proxyml_explain_local / proxyml_explain_local_batch | Feature vector(s) you pass explicitly |

| proxyml_find_counterfactual | One feature vector and a target label |

| proxyml_list_surrogates / proxyml_get_summary / proxyml_export_surrogate / proxyml_diff_models / proxyml_detect_drift | Version IDs and/or threshold values (no feature data) |

| proxyml_get_usage | Nothing |

The typical workflow is designed so that raw training data never leaves your

environment. proxyml_infer_schema derives column metadata from your CSV

locally; proxyml_synthesize_data generates samples server-side from that

schema; and proxyml_train_surrogate sends only those synthetic samples

plus your model's predictions for them. Your original dataset is not

transmitted.

If you use proxyml_explain_local, proxyml_predict_batch, or

proxyml_find_counterfactual with real data points rather than synthetic

ones, those feature vectors will be sent to the ProxyML API.

Authentication
Your API key is sent on every request as an X-API-KEY HTTP header over TLS.

It is read from the PROXYML_API_KEY environment variable and is never

logged or written to disk by this server.

Tools
Schema
| Tool | Description |

|---|---|

| proxyml_infer_schema | Infer a feature schema from a local CSV file — no data sent to the server |

| proxyml_get_schema | Retrieve a stored schema by name |

| proxyml_put_schema | Upload or replace a feature schema |

Training
| Tool | Description |

|---|---|

| proxyml_synthesize_data | Generate synthetic samples from the stored schema |

| proxyml_train_surrogate | Train a linear surrogate on samples scored by your model |

| proxyml_list_surrogates | List trained surrogate models, newest first |

| proxyml_predict_batch | Get surrogate predictions for a list of instances |

Explainability
| Tool | Description |

|---|---|

| proxyml_get_summary | Feature importances and model summary |

| proxyml_export_surrogate | Full coefficient export for audit and governance |

| proxyml_explain_local | Per-feature contribution breakdown for a single instance |

| proxyml_explain_local_batch | Per-feature contributions for multiple instances in one call |

| proxyml_find_counterfactual | Find the nearest point that flips the prediction |

| proxyml_diff_models | Compare feature importances between two surrogate versions |

CI/CD
| Tool | Description |

|---|---|

| proxyml_detect_drift | Compare two versions and return a structured pass/fail against coefficient and fidelity thresholds |

Account
| Tool | Description |

|---|---|

| proxyml_get_usage | Current tier, request count, and quota — useful as a pre-flight check |

Typical workflow

Steps 1–2 are one-time setup. Steps 3–5 can be repeated to retrain as your model changes; use proxyml_diff_models to compare versions.

Agentic workflows
Drift detection in CI/CD
proxyml_detect_drift is designed for use in deployment pipelines. It wraps proxyml_diff_models and applies thresholds to produce a structured pass/fail:

Thresholds can be tuned per use case:

Dev model validation without production data

Validate a model trained in a lower environment by comparing its predictions

against a surrogate trained on production data — no production data required

in the dev environment.
This workflow requires a step the MCP server can't do on its own (scoring

with your dev model), but works naturally in Claude Code where the agent can

execute code directly:

Example prompt for Claude Code:

The surrogate acts as a proxy for production behaviour — if the dev model

agrees with it within tolerance, it's likely behaving consistently with what

was trained on real data.

Counterfactual investigation
When a model makes a decision that needs explaining — a rejected loan application,

a flagged transaction, a declined insurance quote — chain proxyml_explain_local

and proxyml_find_counterfactual to answer both "why?" and "what would need to

change?":

Example prompt:

Claude will call proxyml_explain_local to surface the top contributing features,

then proxyml_find_counterfactual with the target outcome, and interpret the

difference in plain language.

Iterative surrogate improvement
When proxyml_train_surrogate returns a low fidelity warning or other training

diagnostic, the agent can use it to guide the next iteration rather than stopping:

Example prompt:

The training warnings (convergence, sparsity, class imbalance, high correlation)

are designed to be actionable — the agent can read them and decide whether to

adjust num_samples, revisit the schema, or flag for human review.

Governance report
Claude can generate a governance report from existing tools without a dedicated endpoint. Example prompt: