# OpenClaw MCP: how to add MCP servers, and the 10 best to install in 2026

OpenClaw is the open-source personal agent that lives in your messaging apps. You talk to it on WhatsApp, Telegram, Discord, or Slack, and it runs on your own machine with access to your files, your schedule, and your tools. That reach is exactly why the Model Context Protocol (MCP)[Model Context Protocol (MCP)](/articles/what-is-mcp) matters here: MCP servers are how you hand OpenClaw new capabilities without writing an integration yourself.

This guide covers how MCP support works in OpenClaw, the servers that earn a place in its config, and the cases where OpenClaw's own skills system is the better tool. One disclosure up front: we make Parallel, and our search server is our first pick. We'll show the numbers and the tradeoffs so you can judge it on the merits.

## You may already be running the best web search server

Before you install anything, know that OpenClaw has a free, LLM-optimized web search provider powered by the Parallel Search API[free, LLM-optimized web search provider powered by the Parallel Search API](/blog/free-web-search-openclaw). It needs no API key or paid plan, but it isn’t on by default: install the plugin with `openclaw plugins install @openclaw/parallel-plugin`, then set `tools.web.search.provider` to `parallel-free`. OpenClaw never auto-selects a key-free provider, so without that setting it falls back to the first keyed provider it finds, starting with Brave.

So the direct answer to "what's the best web search MCP for OpenClaw" is that the plugin covers basic search for free. Adding Parallel MCP[Parallel MCP](/mcp) on top still makes sense when you want more than search: web_fetch for reading specific pages as an explicit tool, web_decisions for typed questions like a probability or a pick among options, or, with an API key, web_research for cited answers to harder questions and higher rate limits.

## How MCP works in OpenClaw

OpenClaw stores MCP server definitions in `~/.openclaw/openclaw.json` under `mcp.servers`, and the openclaw mcp CLI[openclaw mcp CLI](https://docs.openclaw.ai/cli/mcp) manages them: `openclaw mcp add` builds a definition from flags, `openclaw mcp set` saves one as raw JSON, `openclaw mcp list` and `openclaw mcp show` print what’s saved, and `openclaw mcp status --verbose` reports state without starting any servers. If you prefer a browser, **Settings > MCP** in the Control UI edits the same entries.

Three details trip people up. First, set "transport": "streamable-http" for Streamable HTTP servers, because OpenClaw defaults to SSE when it's omitted, and some hosted servers (Zapier's, for one) don't accept SSE. Second, the CLI commands that change or connect to saved servers (add, set, configure, tools, unset, login, probe, and doctor --probe) need the local Gateway stopped; while it runs, use the Control UI for live changes. Saving proves nothing about reachability, so finish with openclaw mcp doctor <name> --probe, which connects and lists the server's tools. Third, OAuth-protected servers ("auth": "oauth") authenticate through openclaw mcp login, which stores tokens in a shared SQLite store so parallel OpenClaw processes don't fight over refresh tokens.

## The best MCP servers for OpenClaw

### 1. Parallel MCP

Parallel MCP is a hosted server at https://mcp.parallel.ai/mcp. With no account or key it exposes three tools: web_search, which returns ranked results with token-dense excerpts an agent can usually answer from directly, web_fetch, which reads up to 20 URLs per call and returns relevant excerpts or full markdown, PDFs and JavaScript-heavy pages included, and web_decisions, which answers typed questions with sources. An API key, sent as a Bearer header, adds web_research and higher rate limits. It runs on our own web-scale index rather than reselling another engine's results.

OpenClaw's Parallel plugin also provides a search provider (paid, or the key-free Parallel Search (Free) option you select explicitly), so installing the MCP mainly buys you the tools the plugin doesn't have: explicit fetching, web_decisions, and, with an API key from platform.parallel.ai, web_research. Our notes on getting maximum search accuracy out of OpenClaw[getting maximum search accuracy out of OpenClaw](/articles/openclaw-best-practices-web-search) cover the prompting side.

**Best for:** grounding every OpenClaw answer in current web results, plus reading specific pages in depth, at zero cost.

Tradeoffs: we make it, so we're biased. Keyless calls run Search in its low-latency fast mode at lower rate limits; web_research needs an API key and is billed at the underlying API rates. A key also lets you pin a search mode or include and exclude domains through an x-parallel-mcp-config header.

### 2. Granola MCP

If you take meetings in Granola[Granola](https://www.granola.ai/mcp), its official MCP server lets OpenClaw search and read your meeting notes. That pairs well with an agent that lives in your messaging apps: "what did we agree with the supplier on Tuesday?" or "draft the follow-ups from this morning's call" from your phone, answered from structured notes rather than a raw transcript. The server is hosted at https://mcp.granola.ai/mcp over Streamable HTTP and signs you in with OAuth.

Granola says MCP is available on every plan; the free plan covers your own notes from the last 30 days. We haven't run this login end to end in OpenClaw, so confirm it with openclaw mcp doctor granola --probe.

Best for: meeting follow-ups, decisions, and "who said what" questions from chat.

Tradeoffs: it only helps if your meetings already run through Granola, and older notes need a paid plan.

### 3. GitHub MCP

GitHub's official hosted MCP server[official hosted MCP server](https://github.com/github/github-mcp-server) lets OpenClaw read and act on repositories, issues, pull requests, and Actions runs from wherever you're chatting. Asking your agent "what broke in CI overnight?" from your phone is the kind of thing it makes routine.

**Best for:** repo triage, PR review, and CI questions from chat.** Tradeoffs:** it registers a lot of tools, which costs context; limit toolsets to the ones you use, and keep write actions on approval.

### 4. Playwright MCP

Microsoft's Playwright MCP[Playwright MCP](https://github.com/microsoft/playwright-mcp) gives OpenClaw a real browser: navigate, click, fill forms, take screenshots. For a personal agent, that's the fallback for every service that has no API, and by mid-2026 it's among the most-installed MCP servers in the community.

**Best for:** automating sites without APIs and visually verifying what the agent did.** Tradeoffs:** heavier and slower than an API call, and a browser acting on your logged-in sessions deserves sandboxing. For reading pages, `web_fetch` is faster and cheaper; save the browser for interaction.

### 5. Notion MCP

Notion's hosted, OAuth-secured MCP server connects OpenClaw to your notes, docs, and databases. For a personal assistant, that's where captured tasks, reading lists, and meeting notes live, and it turns "add this to my week" messages into actual database rows.

**Best for:** personal knowledge and task workflows.** Tradeoffs:** it writes to your workspace, so keep approval prompts on and scope the connection to the pages the agent actually needs.

### 6. Zapier MCP

Google offers official Workspace MCP servers[Workspace MCP servers](https://developers.google.com/workspace/guides/configure-mcp-servers) for Gmail, Calendar, Drive, and other apps, currently in developer preview and set up through a Google Cloud project with your own OAuth client, which suits people already working in Google Cloud. For a setup that takes a couple of commands, Zapier MCP[Zapier MCP](https://docs.zapier.com/mcp/overview/how-connections-work) is the quicker route. It's a hosted Streamable HTTP server at https://mcp.zapier.com/api/v1/connect that runs actions through the apps already connected to your Zapier account, so “draft a reply to Sam and put Friday’s dinner on my calendar” becomes two tool calls.

In our test, `openclaw mcp login zapier` registered a client and opened Zapier’s authorization page. If OAuth fails in your setup, Zapier’s docs cover a connection token[connection token](https://docs.zapier.com/mcp/get-started/authentication) sent as a Bearer header for clients it doesn’t list.

**Best for:** email, calendar, and the long tail of SaaS apps that have no MCP server of their own.

**Tradeoffs:** each successful tool call uses two tasks[two tasks](https://docs.zapier.com/mcp/features/usage) from your Zapier plan’s allowance, and Zapier sits between your agent and every app, so you’re trusting a second company with those connections. Zapier MCP doesn’t accept SSE, which is one more reason to set the transport explicitly.

### 7. Todoist MCP

Doist runs an official Todoist MCP server[official Todoist MCP server](https://github.com/Doist/todoist-ai) at `https://ai.todoist.net/mcp` over Streamable HTTP with OAuth. It reads and writes tasks, projects, and due dates, which suits the way most people use a messaging agent: firing off “remind me to renew the passport next Tuesday” from a phone and having it land in the right project.

We ran both commands against OpenClaw 2026.10.1 in a sandboxed config; `login` discovered Todoist’s OAuth endpoints and printed the authorization URL with no manual client setup.

**Best for:** capturing tasks from chat and asking “what’s due today?” each morning.

**Tradeoffs:** it has write access to your whole task list, so keep approval prompts on for deletes and bulk edits.

### 8. Home Assistant MCP

Home Assistant ships a Model Context Protocol Server integration[Model Context Protocol Server integration](https://www.home-assistant.io/integrations/mcp_server/) that lets an agent control your devices and read the current state of your home. Because OpenClaw usually runs on a machine inside your network, it can reach `http://homeassistant.local:8123/api/mcp` directly, without the internet exposure that cloud-hosted assistants need. Enable it under **Settings > System > AI**, create a long-lived access token under your profile’s **Security** tab, and keep it in an environment variable.

OpenClaw substitutes `${HA_TOKEN}` when it loads config, reading the process environment or `~/.openclaw/.env`, so the token stays out of `openclaw.json`. Home Assistant also supports OAuth 2.0 with PKCE if you’d rather sign in than manage a token.

**Best for:** “turn off the downstairs lights” and “is the garage door open?” from anywhere you can message OpenClaw.

**Tradeoffs:** the integration requires an administrator account by default and exposes whatever its LLM API can control, so scope the exposed entities in Home Assistant before you connect it. It supports tools, prompts, and the Assist context resource, but not sampling or notifications.

### 9. Linear MCP

Linear’s remote MCP server[Linear’s remote MCP server](https://linear.app/docs/mcp) at `https://mcp.linear.app/mcp` uses Streamable HTTP and OAuth 2.1 with dynamic client registration, and it also accepts a Linear API key as a Bearer token. Its tools find, create, and update issues, projects, and comments, so you can file a bug from a Slack thread or ask what’s blocking this cycle without opening the app.

**Best for:** people who already run their work or side projects in Linear and want to triage from chat.

**Tradeoffs:** it overlaps with GitHub Issues if you use both; pick the one that holds your real backlog, or the agent will guess wrong about where a ticket belongs.

### 10. Context7

Context7[Context7](https://github.com/upstash/context7), from Upstash, serves current documentation and code examples for libraries and frameworks. The hosted server at `https://mcp.context7.com/mcp` works without a key; a free key from the Context7 dashboard, sent as `Authorization: Bearer`, raises the rate limits. It’s the server OpenClaw’s own docs use in their examples, and when we probed it from OpenClaw it reported two tools, `resolve-library-id` and `query-docs`.

**Best for:** coding questions from your phone, where a stale API answer from training data would send you down the wrong path.

Tradeoffs: it covers library docs only; for anything else on the web, Parallel MCP is the better tool, and the two don't overlap much.

## The picks side by side

## When a skill beats an MCP server

OpenClaw grew out of the Pi harness, and it inherited Pi's shell-first instincts: for tools that are really just CLIs, a ClawHub skill is often lighter than an MCP server, because the agent shells out on demand instead of carrying tool schemas in context on every turn. Parallel works both ways; the Parallel CLI ships as an OpenClaw skill[Parallel CLI ships as an OpenClaw skill](https://docs.parallel.ai/integrations/clawhub) if you'd rather skip MCP entirely.

As a rule of thumb, hosted services with OAuth and state (GitHub, Notion) fit MCP; local, composable, context-cheap tooling fits skills. And whichever route you pick, keep the total small. Every server adds tool definitions to the context window and another place for tool-name collisions, and three to five well-chosen servers reliably beat fifteen.

## Frequently asked questions

**Is MCP the best way to extend OpenClaw?** It's one of two good ways. MCP is the right fit for hosted, authenticated services; skills are the right fit for local CLI tooling. OpenClaw can also run as an MCP server itself via `openclaw mcp serve`, so other agents like Claude Code can read and send your channel conversations.

Do I need to configure web search at all? Once. OpenClaw uses a key-free provider only when you select it, so install the Parallel plugin and set tools.web.search.provider to parallel-free for free Parallel search with no API key. Add Parallel MCP when you want page fetching and web_decisions as explicit tools, or an API key for web_research and higher rate limits.

Are these servers free? Parallel MCP's keyless tools, GitHub, Playwright, Home Assistant, and Context7 cost nothing. Parallel's web_research is billed per request at API rates, and a new account comes with up to $80 in free credits plus $5 a month with a card on file. Granola, Notion, Todoist, and Linear work with your existing account, and each successful Zapier MCP call uses two tasks from your Zapier plan.

**Why don’t new tools show up after I save a server?** Saving only writes config. Run `openclaw mcp doctor <name> --probe` to confirm the server connects and lists its tools, then start the Gateway again (CLI edits need it stopped) and send a new message; with hot reload on, changed servers are picked up on the next turn.

## Start with search

The one-line openclaw mcp set command above is the whole setup, and it never asks for an account, a key, or a card. If OpenClaw becomes the way you work, create a free Parallel account[create a free Parallel account](https://platform.parallel.ai/) for web_research and higher rate limits. Then test it on your own questions; the Parallel MCP page[Parallel MCP page](/mcp) has setup for your other clients too.