Best MCP Servers for OpenClaw in 2026: 6 Top Picks by Use Case

The best MCP servers for OpenClaw add a capability OpenClaw doesn't already ship, expose a small set of tools, and pass openclaw mcp doctor --probe. For most setups, that shortlist is Olostep for live web data, the GitHub MCP Server, Context7, Composio, Playwright MCP, and the Filesystem reference server.

An MCP server is a program that exposes tools, resources, and prompts to an AI agent over the Model Context Protocol. If you are new to the protocol, start with how the Model Context Protocol works.

OpenClaw acts as a native MCP client. The OpenClaw MCP documentation states: "Server definitions live under mcp.servers in config, and the tools they expose go through the same tool-profile and tool-policy controls as everything else — connecting a server does not bypass your policy."

Curation matters because the catalog is large. In December 2025, the MCP maintainers reported "Over 97 million monthly SDK downloads, 10,000 active servers and first-class client support across major AI platforms." Most of those servers don't belong in a single always-on OpenClaw agent.

Check What OpenClaw Already Does Before Adding a Server

OpenClaw already ships web search, URL fetching, a managed browser, and memory. Adding an MCP server that duplicates a built-in tool adds tool definitions to the agent's context without adding a new capability.

The OpenClaw docs describe web_search as "a lightweight HTTP tool, not browser automation." web_fetch handles lightweight URL fetching, and the managed browser handles JavaScript-heavy pages and logins.

If you are deciding between OpenClaw's built-in browser and an external one, our guide to browser automation for AI agents covers the trade-offs.

How to Choose MCP Servers for an Always-On Agent

OpenClaw runs on your own machine and takes instructions from chat apps, often while no one is watching. That changes the selection criteria compared with an IDE agent you watch in real time.

Use the four criteria below to score any server before you install it.

Capability Gap

A server earns a slot only when it does a job OpenClaw's built-in tools can't do reliably. GitHub write access, current library docs, OAuth-connected SaaS apps, and large-scale web extraction pass this test. A second search tool usually doesn't.

Some jobs don't need MCP at all. A cron script calling a REST endpoint can be simpler and cheaper, as we explain in MCP versus direct API calls.

Tool Count and Context Cost

Every tool definition costs tokens before the agent reads your message. Anthropic notes that "Most MCP clients load all tool definitions upfront directly into context."

In one illustrative Google Drive-to-Salesforce workflow, Anthropic's engineering team reported: "This reduces the token usage from 150,000 tokens to 2,000 tokens—a time and cost saving of 98.7%." That figure comes from presenting tools as code APIs in a single example, not from a typical OpenClaw setup.

OpenClaw reduces this cost with Tool Search, which defers tool schemas by default in embedded runs and is labeled experimental. The OpenClaw Tool Search docs add a caveat: "Compaction is not always cheaper: small catalogs can gain schema overhead, and additional discovery turns can offset initial payload savings." The dependable fix is to expose fewer tools with toolFilter.include.

Transport and Hosting

OpenClaw supports three transports: stdio, SSE, and Streamable HTTP. The transport decides what runs on your gateway host.

• Stdio: OpenClaw spawns the server as a child process. It needs the server's runtime, such as Node.js or Python, installed on the host.
• Streamable HTTP: OpenClaw connects to a hosted URL. Nothing is installed locally, which suits a headless gateway.
• Auth: Remote servers use either a Bearer token in headers or auth: "oauth" with openclaw mcp login <name>.

Maintainer and Security Posture

Prefer servers maintained by the vendor whose API they wrap. The official MCP servers repository describes its reference servers as educational examples, "not as production-ready solutions," so treat them as starting points.

Check three things before installing: who publishes the server, what token scope it needs, and whether it can write or delete data.

The 6 Best MCP Servers for OpenClaw by Job

Each pick below fills a gap OpenClaw's built-in tools leave open. The table gives the setup facts, and each section explains when to use the server and when to skip it.

Olostep: Live Web Data, Scraping, and Crawling

The Olostep MCP server gives OpenClaw 10 tools for the live web. It turns JavaScript-rendered pages into Markdown, HTML, JSON, or text, so the gateway host doesn't need to run a browser.

• Single pages: scrape_website and get_webpage_content return clean page content, with wait_before_scraping for JavaScript-heavy sites.
• Search and answers: search_web returns structured results, and answers returns a web-grounded answer with sources.
• Site discovery: create_map and get_website_urls list URLs on a site without scraping them.
• Volume jobs: batch_scrape_urls and create_crawl run asynchronously and return an ID, which fits OpenClaw cron jobs that collect results later.

Per the Olostep MCP setup docs, which list all 10 tools, the hosted endpoint is https://mcp.olostep.com/mcp with an Authorization: Bearer header. A local stdio install runs npx -y olostep-mcp with OLOSTEP_API_KEY set.

Olostep does not cover GitHub, local files, or SaaS app actions, so pair it with the servers below. For other search-focused options, see our web search MCP comparison.

GitHub MCP Server: Repos, Issues, and Pull Requests

GitHub's official server lets OpenClaw read code, open issues, and review pull requests. The remote endpoint is https://api.githubcopilot.com/mcp/, so nothing runs on your host.

The full server exposes many tools, so scope it. The GitHub MCP Server README explains: "Enabling only the toolsets that you need can help the LLM with tool choice and reduce the context size." Use a fine-grained token limited to the repositories the agent should touch.

Context7: Current Library Documentation

Context7 returns version-specific documentation and code examples for libraries. It helps when OpenClaw writes or fixes code and would otherwise rely on outdated training data.

The server URL is https://mcp.context7.com/mcp, with your API key passed as an Authorization: Bearer header. Skip it if your agent never touches code.

Composio: SaaS Apps Through One OAuth Endpoint

Composio provides one managed MCP endpoint for 1,500+ tools, including Gmail, Calendar, Slack, and Notion. It handles each app's OAuth sign-in, which saves you from registering OAuth clients yourself.

Connect only the apps your agent needs. Every connected app adds tools, and every OAuth grant adds an account the agent can act on.

Playwright MCP: Scripted Browser Control

Microsoft's Playwright MCP controls a browser through structured accessibility snapshots instead of screenshots. It runs locally with npx @playwright/mcp@latest.

OpenClaw already includes a managed browser, so add Playwright MCP only when you need Playwright-specific behavior. It also requires a browser on the gateway host, which a minimal VPS may not have.

Filesystem: Scoped Local File Access

The Filesystem reference server restricts operations to the directories you pass at startup (or that your client grants through MCP Roots). That makes it useful for giving one agent access to a single folder, such as a notes or exports directory.

It is a reference implementation, so pin the version and keep the allowed directory narrow.

How to Add an MCP Server to OpenClaw

OpenClaw offers three ways to add a server: Settings → MCP in the Control UI, + → Connectors → Add MCP server in the chat composer, or the openclaw mcp CLI. All three write the same mcp.servers entry.

Add a Server From the CLI

The CLI handles both local and remote servers. This example adds Olostep as a remote Streamable HTTP server and exposes only three tools:

openclaw mcp add olostep \
  --url https://mcp.olostep.com/mcp \
  --transport streamable-http \
  --include 'scrape_website,search_web,answers'

A local stdio server uses --command and --arg instead. This example assumes the Filesystem server is installed globally, which avoids npx cold starts:

openclaw mcp add files \
  --command mcp-server-filesystem \
  --arg /srv/openclaw/notes

Add the Bearer header for Olostep through the config or the Settings editor, as shown next.

Write the Config Directly

OpenClaw reads MCP servers from the mcp.servers key in its config. This JSON5 example adds Olostep and GitHub, pulls secrets from environment variables, and filters tools:

{
  mcp: {
    servers: {
      olostep: {
        url: "https://mcp.olostep.com/mcp",
        transport: "streamable-http",
        headers: {
          Authorization: "Bearer ${OLOSTEP_API_KEY}",
        },
        toolFilter: {
          include: ["scrape_website", "search_web", "answers", "create_crawl", "get_crawl_results"],
        },
      },
      github: {
        url: "https://api.githubcopilot.com/mcp/",
        transport: "streamable-http",
        auth: "oauth",
      },
    },
  },
}

• Secrets: The OpenClaw docs say to "Keep credentials out of config literals." ${VAR} substitution or OpenClaw's secret providers keep keys out of the file, and openclaw mcp doctor warns when a header contains a literal value.
• Pausing a server: enabled: false keeps the definition without connecting it.
• Other syntax: Some third-party guides use an mcpServers key or a clawdbot.json file. Neither appears in the current OpenClaw docs, which use mcp.servers in ~/.openclaw/openclaw.json.

Verify the Server Responds

Run a live check after saving:

openclaw mcp doctor olostep --probe

The OpenClaw docs put it plainly: "Saving a definition proves nothing about reachability — the check does." The probe opens a connection and lists the tools the server advertises.

• Zero tools: Check the API key and your toolFilter.include list.
• Stdio won't start: Confirm the command resolves in the Gateway's environment and that cwd exists.
• Changes not visible: Run openclaw mcp reload in the process that owns the connection, or restart the Gateway.

Skills, mcporter, and Native MCP: Which to Use

OpenClaw gives you three ways to extend an agent, and they solve different problems. Skills teach the agent how to work, while MCP servers give it new tools.

The OpenClaw MCP registry docs note that openclaw mcp commands "do not read mcporter servers from config/mcporter.json." For production use, native MCP keeps every tool under one policy and one doctor check. To learn how skills fit alongside servers, read our guide to OpenClaw skills.

Security Rules for MCP Servers in OpenClaw

An always-on agent that reads chat messages and web pages faces more untrusted input than an IDE agent. Every tool you add widens what that input can trigger.

The OpenClaw ecosystem has already seen supply-chain attacks. A Koi Security audit reported in February 2026: "A security audit of 2,857 skills on ClawHub has found 341 malicious skills across multiple campaigns, according to new findings from Koi Security." Those were ClawHub skills, not MCP servers, but the lesson applies to anything you install.

MCP servers carry their own risk. A study of 1,899 open-source MCP servers found that "7.2% of servers contain general vulnerabilities, and 5.5% exhibit MCP-specific tool poisoning."

• Least privilege: Use fine-grained, read-only tokens where the job allows it.
• Allow-lists: Set toolFilter.include so write and delete tools stay off unless needed.
• Secrets: Store keys in environment variables or a secret provider, never in committed config.
• Untrusted content: Treat scraped pages and inbound messages as data, not instructions.
• Approvals: In sessions that support MCP tool approvals, openclaw mcp configure <server> --approval prompt makes the agent ask before using a server that can change state.

A Starter MCP Stack for OpenClaw

Start with two or three servers matched to the agent's job, then add more only when a task fails for lack of a tool. These profiles are a practical starting point:

Run openclaw mcp doctor --probe after each addition and watch whether replies get slower or less accurate. If you also use MCP in other clients, our client-agnostic MCP server roundup covers servers beyond OpenClaw.

FAQ

Does OpenClaw Support MCP Servers Natively?

Yes. OpenClaw is a native MCP client that reads servers from mcp.servers and supports stdio, SSE, and Streamable HTTP transports.

Where Does OpenClaw Store MCP Server Config?

OpenClaw stores MCP servers under the mcp.servers key in its JSON5 config file, ~/.openclaw/openclaw.json by default. The Settings UI and openclaw mcp add write to the same key.

How Many MCP Servers Should I Run in OpenClaw?

Start with two or three servers that fill real gaps. Each server adds tool definitions, so fewer scoped servers usually give the agent clearer tool choices.

What Is the Difference Between mcporter and openclaw mcp?

mcporter is a skill that calls MCP servers through shell commands and keeps its own registry. openclaw mcp manages native servers whose tools pass through OpenClaw's tool policy.

Can OpenClaw Use Remote MCP Servers With OAuth?

Yes. Set auth: "oauth" on an HTTP server and run openclaw mcp login <name>, or sign in from Settings → MCP.

Why Does My MCP Server Show Zero Tools?

The API key is usually wrong or toolFilter.include excludes every tool. Run openclaw mcp doctor <name> --probe to see what the server actually advertises.

Can OpenClaw Act as an MCP Server?

Yes. openclaw mcp serve exposes OpenClaw channel conversations to other MCP clients, which is separate from connecting servers to OpenClaw.