EasyDeploy AI MCP (easydeploy-ai-mcp)

A Model Context Protocol (MCP) server that exposes the EasyDeploy public REST API as tools for Claude, Cursor, Claude Code, and other MCP clients.

PyPI package name: easydeploy-ai-mcp · Import package: easydeploy_ai_mcp

Contents

• Connect with Claude
• What you get
• Requirements
• Install
• Environment variables
• Local MCP (stdio)
• Remote MCP (HTTP)
• Documentation
• REST API reference
• Releases and API compatibility
• Development
• Security
• License

Connect with Claude

Hosted connector (claude.ai or Claude Desktop)

We host the MCP endpoint. You add it once inside Claude; after you connect and sign in, new chats can use EasyDeploy like any other enabled connector.

1. Open Claude in the browser or open Claude Desktop.
2. Open Settings, then Connectors, and choose Add custom connector.
3. Enter exactly:
   • Name: EasyDeploy AI
   • Remote MCP server URL: https://mcp.easydeploy.ai/mcp
4. Save the connector so EasyDeploy appears in your list of connectors.
5. Open the EasyDeploy connector entry and choose Connect, then finish sign-in in your browser. That step authorizes Claude to use your EasyDeploy account.

Connect and sign in (inside Claude): After the custom connector is set up, Claude shows an EasyDeploy card with the MCP URL and a Connect button. Use that flow to sign in. You do not paste an API key into Claude. Access stays tied to the EasyDeploy profile you authenticate in the browser.

Claude Desktop and file uploads: File uploads and some tool calls reach EasyDeploy over the network. On Desktop, Claude blocks outbound traffic unless you allow the domains it should call. Open Settings → Capabilities, turn on Allow network egress, and under the domain allowlist add this one host exactly:

api.easydeploy.ai

That is the only host uploads to the public production server go to; allow it exactly, nothing wider. A server pointed at another EasyDeploy API (one set with EDA_API_BASE) needs that API's host instead, and the start_upload runbook names the exact host to allow.

The domain allowlist UI is available on paid Claude plans.

More detail and variants (for example a URL from your own deployment) are in docs/claude-getting-started.md.

Start here: get_started

get_started returns the EasyDeploy data-science playbook, which is most useful at the start of a modeling task. Its description and the server's MCP instructions say what it returns rather than when the agent must call it, as the Claude connector directory requires. Hosts cache tool descriptions and server instructions, and many never show the instructions to the model; a tool result is fetched live on every call. So the playbook lives in the tool's result, not in its description: the name and description stay fixed, and the content is updated with each release without anyone reconnecting.

• get_started() returns the overview: the roles, the whole workflow with the tools at each step, and the hard rules.
• get_started(section=...) returns one section: prepare, split, upload, train, validate or predict. all returns every section.
• Each result carries guide_version, updated, the list of sections with a one-line summary each, and next, the section to read next.

The content ships inside the package as Markdown, one file per section, in src/easydeploy_ai_mcp/guide/. The tool is read-only and makes no API call.

You are the data scientist

EasyDeploy searches models, engineers features, tunes hyperparameters, trains, deploys and predicts. It does not clean your data, define the target, split it, or check the model on data it never saw. The agent does that, following the get_started playbook:

1. Prepare. Agree the decision and the target with you, explore the data, drop identifiers and anything not known at prediction time from the training file, and write a one-sentence target definition. Ids can stay in test and scoring files: prediction ignores columns the model wasn't trained on and returns them in the output, which makes joining results back easy. Data changes and dropped columns are confirmed with you before they are applied.
2. Split. Stratified 80/20 for classification, chronological for time series, no entity in both files, no duplicates. Balancing such as SMOTE touches the train file only.
3. Upload and train. The files go up separately, with dataset_type train and test (validation optional). The model version is created on the train dataset version only.
4. Validate on the holdout. The report's cross-validation score is measured inside the training file and its training-fit metrics are in-sample. The real estimate comes from run_batch_prediction on the test dataset version: the downloaded CSV keeps every input column, the true label included, in the original row order and adds prediction and probability_<class> columns. The agent computes the metrics from it and picks the decision threshold from the cost of each kind of error (best F1 only when those costs are unknown), or reports the error margin for regressors.
5. Use real outputs. Dashboards, reports and scored lists are built from real predictions, never mock values.

Getting a training file into EasyDeploy

start_upload opens an upload session and the agent picks the best byte channel the host offers. File contents never travel through the conversation or through tool arguments on any channel — the agent passes a URL or runs a curl, never the data itself.

get_upload_status reports URL_ISSUED → RECEIVING → UPLOADED → VALIDATING → READY | REJECTED → CONSUMED (plus EXPIRED). complete_upload requires READY — every byte that becomes a dataset version passes the CSV validator first, and a session still in flight, REJECTED or EXPIRED returns a 400 naming what to do. On REJECTED, the error field says which check failed. Retrying complete_upload is safe: on a session that is already CONSUMED it returns the existing dataset version, and a 409 means another call for the same session is still in progress (wait a few seconds and retry). get_upload_status only covers sessions opened by start_upload; files uploaded through the web model builder are registered in the web app and do not appear there.

New dataset or new version. Omit dataset_id on start_upload and complete_upload creates a new dataset with the name you give it. Pass an existing dataset's id and the upload becomes a new version of that dataset; complete_upload's name is then ignored and the dataset keeps its name. complete_upload returns dataset and datasetVersion once each, with the column names as a columnNames list. qa_status is informational and does not gate anything. version_type matters in one case: submit_training_job with an explicit dataset_version_id rejects a raw version. Omit dataset_version_id to train on the model version's own dataset.

Fallback when the host has no byte channel at all (no egress and no file bridge): the agent saves the training file, hands it to you as a download, and points you at the web model builder using the fallback text start_upload returns, which carries the right model-builder URL for the environment (for example easydeploy.ai/model-builder). After you upload there, tell the agent the dataset name you entered or paste the dataset URL from the page — it resolves the dataset with list_datasets and confirms the row count before continuing. That upload is already registered as a dataset, so complete_upload is not used for it. Adding the allowlist entry above (where your plan allows it) avoids the detour.

Reading a model report

get_model_report returns the report plus a structured metrics object (passed through unchanged from the API). Model search scores candidates with 10-fold cross-validation (stratified and shuffled on ROC-AUC for classifiers, shuffled on negative MSE for regressors; forward-chaining TimeSeriesSplit for a version created with time_series_mode=true), then refits the winner on all rows. There is no separate test set: accuracy, the confusion matrix, per-class precision/recall and in-sample ROC-AUC under metrics.trainingFit are training fit, and metrics.crossValidation.score is the report's out-of-sample estimate, measured inside the training file. The holdout estimate comes from scoring your own test file with run_batch_prediction, not from the report. metrics.crossValidation.strategy records which CV a run used (time_series_split for forward-chaining), and metrics.warnings flags anything to be careful quoting. The prose summary is written by an LLM and may use looser wording.

Local MCP on your computer (stdio)

Run the MCP server on your own machine using your EasyDeploy API key. Nothing is exposed to the internet.

1. Install

pip install easydeploy-ai-mcp

2. Add to Claude Desktop config

Edit (or create) the Claude Desktop config file:

Merge the following into the root of that JSON (keep any existing keys):

{
  "mcpServers": {
    "EasyDeploy AI": {
      "command": "easydeploy-ai-mcp-stdio",
      "env": {
        "EDA_API_KEY": "eda_live_YOUR_KEY"
      }
    }
  }
}

Replace eda_live_YOUR_KEY with your key from Account → API Keys in the EasyDeploy dashboard. Use the full path to easydeploy-ai-mcp-stdio (run which easydeploy-ai-mcp-stdio to find it) if Claude cannot locate it on your PATH.

3. Restart Claude Desktop

Fully quit and reopen the app. EasyDeploy AI will appear in your MCP servers.

For self-hosting on Docker or a cloud provider, see Remote MCP (HTTP).

What you get

• 28 tools: get_started (the playbook) plus tools covering projects, datasets (including the three upload channels), model versions, training jobs, predictions, and account status.
• A title and tool annotations on every tool (title, set both on the tool and as annotations.title, plus readOnlyHint, destructiveHint, idempotentHint, openWorldHint), so hosts can tell reads from writes. get_started and the 16 other get_* / list_* tools are read-only. A write is destructiveHint: true when it can overwrite fields of an existing record: create_project, create_model and create_dataset_version (when given an id they update that record), update_dataset, and submit_training_job (a resubmitted job retrains its model version in place). The other writes only add records and are destructiveHint: false; no tool deletes anything. Only upload_from_url and start_upload (given a host file) make EasyDeploy fetch an external URL. create_project and create_model are create-or-update, so they are marked as non-idempotent writes. Renaming a dataset is its own tool, update_dataset; get_dataset is a pure read.
• stdio transport for local clients, or HTTP with Streamable MCP on /mcp and GET /healthz for load balancers.
• Hardening: HTTPS-only calls to the EasyDeploy API; optional MCP_SERVICE_TOKEN for the HTTP MCP surface; response fields trimmed where appropriate for agents; query values (OAuth codes, state, PKCE values) left out of the access log; Cache-Control: no-store on token and registration responses.

For production and SOC 2–sensitive setups, prefer self-hosting so your data stays within your own infrastructure. The hosted connector at https://mcp.easydeploy.ai/mcp is fine for most users.

Requirements

• Python 3.10+
• An EasyDeploy API key from the dashboard (Account → API Keys). The client uses the production EasyDeploy API host by default.

Install

From PyPI (after first release)

pip install easydeploy-ai-mcp

From source

git clone https://github.com/easydeploy-ai/easydeploy-ai-mcp.git
cd easydeploy-ai-mcp
pip install -e ".[dev]"   # includes pytest
# or minimal runtime only:
pip install .

Environment variables

Local MCP (stdio)

Use when the client starts the server as a subprocess (Claude Desktop, Cursor, etc.).

export EDA_API_KEY="eda_live_..."
easydeploy-ai-mcp-stdio

Or: python -m easydeploy_ai_mcp

Example config snippet:

{
  "mcpServers": {
    "easydeploy-ai": {
      "command": "easydeploy-ai-mcp-stdio",
      "env": {
        "EDA_API_KEY": "eda_live_..."
      }
    }
  }
}

Remote MCP (HTTP)

Serves Streamable HTTP via FastMCP on /mcp (confirm with your pinned FastMCP 3.x version). Health checks: GET /healthz.

export EDA_API_KEY="eda_live_..."
easydeploy-ai-mcp-http

Or: uvicorn easydeploy_ai_mcp.http_main:app --host 0.0.0.0 --port 8080

If you embed mcp.http_app() in another ASGI app, pass through lifespan from the FastMCP HTTP app (FastMCP ASGI); easydeploy_ai_mcp.http_main already does this for uvicorn.

Auth modes

Pick exactly one (setting both EDA_OAUTH_ENABLED and MCP_SERVICE_TOKEN raises at import):

• OAuth 2.0 resource server (multi-tenant): set EDA_OAUTH_ENABLED=1 plus EDA_COGNITO_USER_POOL_ID and EDA_COGNITO_CLIENT_ID. Install the optional extra: pip install easydeploy-ai-mcp[oauth]. The server validates incoming Cognito access JWTs locally against the Cognito JWKS (issuer, signature, exp, token_use=='access', client_id) and forwards the token to the EasyDeploy API. EasyDeploy API keys (prefix eda_live_) are accepted in the same Authorization: Bearer header and forwarded as-is — the API is the source of truth for revocation. RFC 9728 metadata is published at /.well-known/oauth-protected-resource; RFC 8414 proxy metadata includes registration_endpoint and grant_types_supported (authorization_code, refresh_token), and POST /oauth/register returns the static Cognito MCP EDA_COGNITO_CLIENT_ID (RFC 7591-style, public client), with the requested grant_types it supports echoed back (authorization_code and refresh_token; both when none are requested). 401 responses include WWW-Authenticate: Bearer … so MCP clients can discover the auth server. Note: Cognito access tokens carry client_id, not aud; do not configure an audience.
• Shared-secret gate (legacy single-tenant): set MCP_SERVICE_TOKEN. All outbound API calls use the static EDA_API_KEY.
• No auth: development only.

Brokered authorization

Cognito matches redirect_uri literally against the app client's callback list — no wildcards, no prefixes, no ports — and it checks at /oauth2/authorize, before rendering the Hosted UI. A client whose callback is not registered therefore never sees a sign-in page: the browser lands on redirect_mismatch, which reads as "the connector could not load the login page" rather than as a configuration problem.

That works for clients with one fixed callback and cannot work for the rest:

Set EDA_MCP_OAUTH_BROKER=1 and the server sends Cognito its own callback, {issuer}/oauth/callback, which is one fixed URL you register once. It seals the client's real redirect_uri into the OAuth state, and when Cognito comes back it forwards the authorization code to that URI with the client's own state restored. /token rewrites redirect_uri to match, because Cognito checks it a second time at the token exchange.

This moves the redirect check from Cognito to this server, so it stays a real check. Without one, anybody could start a flow here with a callback they control and collect a signed-in user's authorization code. oauth_broker.py is about as tight as the list it replaces: an HTTPS callback must match a known client's host exactly and, where that client's callback path is known, the path too (claude.ai/api/mcp/auth_callback, chatgpt.com/connector/oauth/…, vscode.dev/redirect). Subdomains are never implied — one open redirect or dangling CNAME under a vendor's domain would otherwise be a place to collect codes. Loopback is accepted per RFC 8252 §7.3, and private-use schemes from a short allowlist. Extend it with EDA_MCP_EXTRA_REDIRECT_HOSTS / EDA_MCP_EXTRA_REDIRECT_SCHEMES rather than by widening the code.

PKCE is mandatory in broker mode. Cognito treats code_challenge as optional, and the broker's /token rewrite removes the one check Cognito made there, so a flow without PKCE would leave the code bound to nothing — anyone who found it in a browser history or a proxy log could redeem it. /authorize refuses a brokered request without code_challenge_method=S256. Every MCP client already sends one.

Rollout order matters. Register {issuer}/oauth/callback on the Cognito app client first, then set EDA_MCP_OAUTH_BROKER=1. Doing it the other way round breaks the clients that work today, which is why the default is off.

Docker — run locally (same image you deploy to ECS/Fargate; includes easydeploy-ai-mcp[oauth]):

# Convenience: build + run (reads .env in the repo root if present)
./scripts/run_mcp_docker_local.sh

# Explicit env vars (no .env)
./scripts/run_mcp_docker_local.sh -e EDA_API_KEY="eda_live_..."

# Different host port
PORT=9000 ./scripts/run_mcp_docker_local.sh

Host on AWS (Fargate + ALB): build and push this repo’s Dockerfile to a container registry, then deploy behind an HTTPS load balancer. Set the env vars listed above on the task/container.

Manual equivalent:

docker build -t easydeploy-ai-mcp .
docker run --rm -p 8080:8080 \
  -e EDA_API_KEY="eda_live_..." \
  easydeploy-ai-mcp

Documentation

• docs/claude-getting-started.md — EasyDeploy + Claude: Connectors or local Desktop config JSON
• docs/claude.md — Claude Connectors vs Claude Code, transports, headers

REST API reference

This MCP server is a thin client over the EasyDeploy public REST API. Endpoint behavior, request bodies, and response shapes are defined by EasyDeploy (dashboard, product help, and official API materials at easydeploy.ai). This repo does not duplicate the full OpenAPI spec; it maps those operations to MCP tools.

Releases and API compatibility

This repository is the open-source home of the EasyDeploy MCP server. New releases track the EasyDeploy public REST API as documented for customers (dashboard and official API materials). If the API adds or changes endpoints, expect corresponding updates here. Contributors should follow CONTRIBUTING.md when changing tools or client behavior.

Development

pip install -e ".[dev]"
pytest

Optional — real Cognito JWT against the HTTP app (live JWKS, no mocks): set EDA_INTEGRATION_COGNITO_ACCESS_TOKEN plus the same EDA_COGNITO_* vars you use for OAuth mode, then run pytest tests/test_cognito_jwt_integration.py -v. See the docstring in that file.

See CONTRIBUTING.md for pull requests and reporting issues.

Security

See SECURITY.md for vulnerability reporting and deployment notes.

License

MIT — see LICENSE.