Virix LabsHuman Browser by Virix Labs
A2A ProtocolPricingBlogCompareDocs
AccountRefundPrivacyTerms
Get Started →
June 15, 2026 · 9 min read · AI Agents · OpenClaw

OpenClaw and MCP (2026): Connecting Your Personal Assistant to External Tools

OpenClaw Gateway hub connected by lines to multiple MCP server boxes — illustration of MCP integration with OpenClaw

If you searched openclaw mcp, you're probably trying to answer one of two questions. Either "does this support MCP at all?" (yes), or "how do I plug my existing MCP servers into OpenClaw without re-implementing them as skills?" (also yes — with examples). This post covers both, with a clear take on when MCP beats a skill and vice-versa.

Disclosure: we make Human Browser, which ships an MCP server for protected-site browser work. We'll cover the OpenClaw + Human Browser MCP integration specifically; the general MCP material applies to any server.

Contents

  1. MCP, in one paragraph
  2. How OpenClaw uses MCP
  3. Skills vs MCP — when each wins
  4. Wiring an MCP server into OpenClaw
  5. Three servers worth installing first
  6. OpenClaw + Human Browser MCP
  7. FAQ

MCP, in one paragraph

The Model Context Protocol (MCP, shipped by Anthropic in late 2024 and rapidly adopted by everyone else) is a standard for giving an LLM agent typed tools. An MCP server is a process that exposes a set of named tools with JSON schemas — github.create_issue, filesystem.read_file, browser.navigate. An MCP client is the agent — Claude Code, Cursor, ChatGPT Desktop, OpenClaw — that calls those tools. The protocol handles handshake, capabilities, calls, and streaming. The point is that the agent doesn't have to bake in support for every API; it discovers tools at runtime and calls them generically.

Skills are guidance. MCP is capability. Pick the right one for the right job — or, more often, use both.

How OpenClaw uses MCP

OpenClaw plays both roles:

The local MCP mode is most visible on the Windows Hub companion app — there's an explicit tray toggle for it. On macOS and Linux, you enable it via the Gateway config or the --mcp flag when starting the Gateway. Either way, the OpenClaw MCP endpoint becomes available on a loopback port and other MCP clients can connect.

Skills vs MCP — when each wins

This is the question that's actually worth thinking about. The two systems look like alternatives but solve different problems.

Use a skill when you need guidance

The agent already has a tool that can do the job (file system, browser, shell, MCP-mediated API). What it doesn't have is the domain knowledge for this team's PR review style, your calendar conventions, your customer-support tone. That guidance lives in a SKILL.md — Markdown the agent reads as authoritative instructions. The agent uses existing tools; the skill tells it how and when.

Distribution path: ClawHub (openclaw skill add <name>) or local folders.

Use MCP when you need a capability

The agent doesn't have the tool at all. It can't talk to Stripe's API. It can't list issues from your private GitHub Enterprise. It can't drive a stealth browser. MCP gives the agent a typed surface to call into. The protocol-level abstraction means you (or someone else) writes the server once and any MCP-speaking agent can use it.

Distribution path: the MCP server registry, your own internal repos, or a hosted endpoint.

The honest answer

Real production OpenClaw setups use both. A skill says "when the user asks for a PR review, fetch the diff via github.get_pr_diff (an MCP tool) and summarise per our team's style." The skill is the playbook; the MCP server is the toolbox. Picking one over the other is rarely the right framing.

Hub diagram with OpenClaw Gateway at the center connected to three MCP server boxes labelled GitHub MCP, Chrome MCP, and Human Browser MCP

Wiring an MCP server into OpenClaw

Concretely, the Gateway config has an mcp.servers object, keyed by server name. Each entry gives the launch command (or a url for a remote server) and any env vars it needs:

# Open the config (creates one with defaults if missing)
$ openclaw config edit

# In your editor — add an MCP server entry:
{
  "mcp": {
    "servers": {
      "github": {
        "command": "npx",
        "args": ["-y", "@modelcontextprotocol/server-github"],
        "env": {
          "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_..."
        }
      }
    }
  }
}

# Restart the Gateway so the new server is picked up
$ openclaw restart

From the agent's point of view, the tools exposed by that MCP server are now available the same way bundled tools are. Ask it "list my open PRs" and it will call github.list_pull_requests without you wiring the call explicitly.

Three servers worth installing first

From the broader MCP ecosystem, three give immediate, durable value in an OpenClaw setup:

1. GitHub MCP

The agent can read issues, list PRs, post comments, create branches, and merge PRs (with permissions). The official server is @modelcontextprotocol/server-github; the only setup is a fine-grained PAT in env. The first "summarise the failing CI on PR #123" from your phone via WhatsApp justifies it.

2. Chrome MCP (or browser-use)

Attaches the agent to your real signed-in Chrome (via the Chrome MCP extension) or to a fresh Chromium process. Useful for personal-assistant tasks against sites where you're logged in — Gmail summarisation, Calendar triage, internal dashboards. OpenClaw's bundled browser tool is the better choice when you want the isolated openclaw profile; Chrome MCP is for "use my real session."

3. A filesystem MCP for the directories you care about

@modelcontextprotocol/server-filesystem, with the allowed roots restricted to the project directories the agent should touch. Lets the agent grep, summarise, and edit files without inheriting the full power of the shell. Pair with a permission model that requires confirmation for writes.

OpenClaw + Human Browser MCP

Chrome MCP is great for "drive my logged-in Chrome." It is not the right tool when the target site fights bots — Cloudflare interstitials, DataDome challenges, PerimeterX press-and-hold. For that case there's a different MCP server: the Human Browser MCP. It exposes the same browser-shaped tool surface (navigate, click, type, screenshot, evaluate) but the browser on the other end is a stealth-patched Chromium running on a residential IP, with a CAPTCHA solver and human-paced input baked in.

From OpenClaw's side, it's a normal MCP server entry:

{
  "mcp": {
    "servers": {
      "human-browser": {
        "command": "npx",
        "args": ["-y", "@virixlabs/humanbrowser", "mcp"],
        "env": {
          "HB_TOKEN": "hb_live_..."
        }
      }
    }
  }
}

Add a skill that says "for browser tasks against protected sites (Cloudflare, DataDome, PerimeterX), prefer the human-browser.* tools over the local browser.* tools." Now the agent picks the right browser for the target without you thinking about it.

The full integration recipe (including the alternative remote-CDP path and when to pick MCP vs CDP) is in our companion post: OpenClaw browser: the tool, the remote-CDP mode, and when to pair it with Human Browser.

Try the Human Browser MCP — $1 for about 4 typical tasks

Drop the npx @virixlabs/humanbrowser mcp server into your OpenClaw mcp.servers and your agent reaches Cloudflare- and DataDome-protected sites without changing skills or prompts. Pay-as-you-go: $0.10/hr browser time, $4/GB residential proxy, $0.005/captcha, $0.02 per agent step with the AI included. No subscription. humanbrowser.cloud.

FAQ

Does OpenClaw support MCP?

Yes — both as a client (consuming external MCP servers) and as a server (exposing its capabilities to other MCP clients).

Skills vs MCP — which do I use?

Use a skill for guidance (when, how, your team's conventions). Use an MCP server for capability (tools the agent doesn't have). They compose — a skill can tell the agent which MCP tool to call.

How do I add an MCP server?

openclaw config edit, add an entry under mcp.servers with the launch command + env, then openclaw restart.

Which MCP servers should I install first?

GitHub MCP, Chrome MCP (or Human Browser MCP for protected sites), and a scoped filesystem MCP.

Can OpenClaw connect to Human Browser MCP?

Yes — it's an ordinary MCP server entry (npx @virixlabs/humanbrowser mcp). The MCP tools proxy to a stealth-patched Chromium on a residential IP.

Related Articles

Top up — $20

Loading secure checkout…

More coins →