Evomi MCP Server

A Model Context Protocol (MCP) server for Evomi's proxy and web scraping APIs. It gives AI assistants like Claude direct access to a proxy account — credentials, connection strings, geo targeting, usage, session rotation — alongside a full scraping toolkit for crawling sites and extracting structured data.

39 tools by default, 40 with EVOMI_ENABLE_SPENDING=1.

Features

Proxy & account management

• Proxy Credentials - Username and password for each product on the account
• Connection Strings - Geo-targeted, session-pinned strings, with a curl command to check them
• Bulk Proxy Lists - Up to 50 endpoints per call
• Geo Targeting - Search the countries, regions, cities and ISPs a product actually offers
• Usage & Balance - Bandwidth over 24h, 3d and 7d, in total and per bucket
• Session Control - Rotate a sticky session onto a new exit IP
• Browser Profiles - List saved fingerprint profiles, and order more behind an opt-in
• Service Access - Scraper and Browser credits, concurrency and endpoints

Web scraping

• Single Page Scraping - Scrape any URL with automatic JavaScript detection
• Website Crawling - Multi-page crawling with depth control
• URL Discovery - Find URLs via sitemaps, CommonCrawl, or in-site crawling
• Domain Search - Find domains by searching the web
• AI-Powered Extraction - Use AI to extract structured data from pages
• Conversational Agent - Natural language interface for scraping tasks
• Config Management - Save and reuse scraping configurations
• Schema Management - Define and test extraction schemas
• Storage Configuration - Manage cloud storage for scraped data
• Scheduled Jobs - Automate scraping on a schedule

Installation

pip install evomi-mcp

Needs Python 3.10 or newer. The server runs on both major versions of the mcp SDK (>=1.8.0,<3), so it installs into an environment already pinned to 1.x as well as a fresh one that resolves 2.x.

Run evomi-mcp --help for the environment variables it reads, or evomi-mcp --version for the installed version. With no arguments it speaks MCP over stdio, which is how an MCP client starts it.

Configuration

One credential is enough:

export EVOMI_PUBLIC_API_KEY="your-public-api-key"

Take it from Settings > API for a personal account, or Settings > Team for a team one. It authenticates the proxy and account tools directly, and the scraping tools authenticate with the account's scraper key, which the server reads from the Public API on the first scraping call and keeps in memory for the process. It is never written to disk and never appears in tool output.

Where a specific variable is set it wins over EVOMI_API_KEY. Setting a scraper key alone works too, and the proxy and account tools then need EVOMI_PUBLIC_API_KEY as well, since one key cannot serve both APIs.

Optional settings:

export EVOMI_BASE_URL="https://scrape.evomi.com"   # scraper API, default
export EVOMI_PUBLIC_BASE_URL="https://api.evomi.com"  # public API, default
export EVOMI_HIDE_PROXY_PASSWORDS=1  # mask every proxy password and service API key
export EVOMI_ENABLE_SPENDING=1  # register the tools that spend account balance

Credentials in tool output

get_proxy_credentials, build_proxy_connection_string and generate_proxy_list return live proxy passwords, which is what they are for, and get_api_access returns a service API key when asked with include_api_key. Every other tool returns balances, endpoints, usage and targeting data only.

Four things bound that. Those four tools carry an instruction in their descriptions not to repeat the value back unless it was asked for directly — the MCP spec has no annotation for a sensitive result, so the description is the only channel that reaches the model. generate_proxy_list and build_proxy_connection_string return at most 50 entries per call, well under the 500 the Public API allows, and refuse a larger request rather than clamping it. The curl_example that comes with a connection string has its password masked, since it is the field most likely to be pasted into a terminal or a ticket; pass runnable_curl_example: true for the form that can be run. And EVOMI_HIDE_PROXY_PASSWORDS=1 turns disclosure off entirely:

The two that refuse do so because their entire output is the credential: a connection string with a masked password cannot connect. The refusal points at list_proxy_products and list_proxy_targeting_options, which give the gateway hostname, ports and targeting values with no credential in them, and the descriptions the model sees change too.

EVOMI_ENABLE_SPENDING=1 is the opposite, an opt-in for the one tool that costs money. Without it order_browser_profile is not registered at all, so the connected model never sees it.

Tool annotations

Every tool sets all four of the MCP spec's hints. destructiveHint and openWorldHint default to true, so a tool that omits them advertises itself as potentially destructive and as reaching an unbounded external world.

openWorldHint is true for the tools that reach an address the caller chose (scrape_url, crawl_website, map_website, search_domains, agent_request, and the schema and config tools that validate against the page they describe), and false for everything that only talks to Evomi's own endpoints. readOnlyHint is false for the tools that create, update, delete, toggle, rotate or order something. The MUTATING prefix on rotate_proxy_session and order_browser_profile is in the description as well, because the description reaches every model where an annotation only reaches a client that reads it.

Usage with Claude Desktop

Add to your Claude Desktop configuration (~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

{
  "mcpServers": {
    "evomi": {
      "command": "evomi-mcp",
      "env": {
        "EVOMI_PUBLIC_API_KEY": "your-public-api-key"
      }
    }
  }
}

Or if installed from source:

{
  "mcpServers": {
    "evomi": {
      "command": "python",
      "args": ["-m", "evomi_mcp.server"],
      "env": {
        "EVOMI_PUBLIC_API_KEY": "your-public-api-key"
      }
    }
  }
}

Available Tools

Proxy & Account (9 tools, Evomi Public API, + 1 opt-in)

Scraping Operations (6 tools)

Config Management (6 tools)

Schema Management (6 tools)

Storage Management (4 tools)

Schedule Management (7 tools)

Account (1 tool)

Tool Examples

Scraping

// Basic scrape
{"url": "https://example.com"}

// AI extraction
{"url": "https://example.com/products", "ai_enhance": true, "ai_prompt": "Extract product names and prices"}

// Browser mode with actions
{"url": "https://example.com", "mode": "browser", "js_instructions": [{"click": ".accept-cookies"}, {"wait": 1000}]}

Crawling

// Basic crawl
{"domain": "example.com", "max_urls": 50}

// With URL filter
{"domain": "example.com", "url_pattern": "/products/", "depth": 3}

Domain Search

// Find domains
{"query": "best e-commerce sites for electronics", "max_urls": 20, "region": "us-en"}

// Up to 10 queries in one call, max_urls applying to each
{"query": ["online bookstores", "book shops UK"], "max_urls": 50}

Config Management

// Create config
{"name": "Product Scraper", "config": {"mode": "browser", "output": "markdown"}}

// Generate config with AI
{"name": "Amazon Scraper", "prompt": "Scrape product title, price, and reviews from Amazon"}

Scheduling

// Create daily schedule
{"name": "Daily Prices", "config_id": "cfg_abc123", "interval_minutes": 1440, "start_time": "09:00"}

Credits

Scraping operations consume credits, and what a call costs depends on the mode it runs in and the options it carries. The current rates are at docs.evomi.com; get_api_access reports the balance on the account.

Each scraping response carries credits_used and credits_remaining, so the cost of a call is visible in its own result.

Development

Setup

pip install -e ".[dev]"

Running the Server Directly

evomi-mcp
# or
python -m evomi_mcp.server

Tests

pytest

The suite runs against both mcp majors, and CI runs it on each of them across the supported Python versions.

Releasing

The version is declared in both pyproject.toml and src/evomi_mcp/__init__.py, and CI fails if they disagree. Publishing a GitHub Release tagged vX.Y.Z builds the artifacts, runs the suite against the built wheel and uploads to PyPI over Trusted Publishing, so there is no token anywhere.

Links

• Evomi Website
• Evomi Dashboard
• API Documentation

License

MIT — see LICENSE.