Are MCP servers free? Protocol, licenses and service costs

Mostly yes for the software: MCP is an open protocol and most MCP servers, including the reference servers, are open source and cost nothing to run. What can cost money is the service a server connects to (each vendor’s own plan applies, often with a free tier) and the model usage in the client that calls the server.

Applies to: the Model Context Protocol and MCP servers listed in the Aftergradient MCP directory; vendor pricing pages as of October 2, 2026. Last checked: October 2, 2026.

On this page

Three separate costs

The protocol is open

The MCP documentation describes MCP as “an open-source standard for connecting AI applications to external systems”. The specification lives in a public repository, modelcontextprotocol/modelcontextprotocol. Its LICENSE file says the project is moving from the MIT License to Apache-2.0: new code and specification contributions are Apache-2.0, documentation (excluding specifications) is CC-BY-4.0, and older contributions whose authors have not agreed to relicense stay under MIT. None of these licenses charge a fee to implement the protocol.

Most server software is open source

Reference servers

The modelcontextprotocol/servers repository holds a small set of reference servers maintained by the MCP steering group: Everything, Fetch, Filesystem, Git, Memory, Sequential Thinking and Time. The repository carries the same MIT-to-Apache-2.0 licensing note as the specification (LICENSE). They run locally through npx or uvx, for example:

npx -y @modelcontextprotocol/server-memory

uvx mcp-server-git

Warning

The servers README says these are reference implementations meant as educational examples, “not as production-ready solutions”, and that developers should evaluate their own security requirements.

Directory listings: Filesystem, Fetch, Git, Memory, Sequential Thinking, Time.

Vendor servers and their licenses

Licenses as stated in each repository’s LICENSE file:

Sentry’s license is the Functional Source License: it permits any purpose other than a “Competing Use”, which it defines as offering the software to others in a commercial product or service that competes with Sentry, and it names internal use as a permitted purpose (LICENSE.md). Read the full text before you build a product on it.

Some vendors only offer a hosted endpoint without a server repository you can run yourself (Linear, Notion’s current server, Vercel, Figma). Their use falls under the vendor’s account terms rather than an open-source license.

Hosted servers bill through the underlying service

A hosted MCP server acts on your account with the service. Its tools spend whatever that service charges for: API credits, plan seats or paid features. Some examples from official pages:

GitHub

GitHub’s docs say the GitHub MCP server “is available to all GitHub users regardless of plan type”, but tools inherit the access requirements of the matching GitHub feature: if a feature needs a paid GitHub or Copilot license, the MCP tool needs the same (GitHub Docs). Copilot plans are listed in the plans tracker.

Microsoft Learn

Microsoft states there is no charge to use the Microsoft Learn MCP Server and no authentication is required (Microsoft Learn).

Firecrawl

Firecrawl’s pricing page lists a Free plan with 1,000 credits every month, refreshed monthly, with no credit card required. Its MCP server README adds a hosted keyless tier at https://mcp.firecrawl.dev/v2/mcp where scrape, search and parse work without an API key, rate-limited; crawl, map and agent still need a key.

Exa

Exa’s pricing page says Exa is pay-as-you-go with no subscription; new accounts get $20 in free credits and the Free Tier adds $10 in credits every month. The Exa MCP README says the hosted server works anonymously with rate limits, and that OAuth or an API key gives higher limits and access to Exa Agent.

Tavily

Tavily’s pricing page lists a Free plan with 1,000 API credits per month and says new accounts start with those credits with no credit card required. When credits run out, requests stop until the monthly reset or an upgrade.

Context7

Context7’s plans page lists a Free plan at $0 with 1,000 API calls per month for public repositories; private repositories need a paid plan.

For other servers, check the vendor’s own pricing page before connecting a server to production data. A free MCP server does not change what the service bills for the operations it performs.

Model usage in the client

Every tool call happens inside an agent session, and the session uses model tokens. In Claude Code, Anthropic’s costs page says Claude Code charges by API token consumption and points to claude.com/pricing for subscription plans; on Pro and Max, usage is included in the subscription. See the Claude Code plans tracker for plan limits, and the API billing listing for pay-as-you-go.

MCP tools add to that usage because their definitions and results enter the context. Claude Code’s MCP docs say tool search is on by default, so only tool names and server instructions load at session start and full definitions load when needed. The same page says Claude Code warns when an MCP tool’s output exceeds 10,000 tokens and caps it at 25,000 tokens by default, adjustable with MAX_MCP_OUTPUT_TOKENS:

export MAX_MCP_OUTPUT_TOKENS=50000

A higher cap lets larger tool results into context, which uses more of your plan or API budget.