- [The best MCP servers for Codex CLI in 2026](https://parallel.ai/articles/best-mcp-servers-for-codex-cli)
Tags:Guides
Reading time: 5 minAugust 17, 2026
# The best MCP servers for OpenClaw in 2026
OpenClaw supports MCP servers through its gateway config, but its most important web capability ships by default. This guide covers how MCP works in OpenClaw, the five servers worth installing, and when a skill is the better choice.
Tags:Guides
Reading time: 7 minOpenClaw 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)] 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 ships with free, LLM-optimized web search by default, powered by the Parallel Search API[free, LLM-optimized web search by default, powered by the Parallel Search API]. A fresh OpenClaw install can already ground its answers in current web results without any configuration, an API key, or a paid plan.
So the honest version of "what's the best web search MCP for OpenClaw" is: you probably have it already. Adding the Parallel Search MCP on top still makes sense in two cases: you want web page fetching exposed as an explicit tool alongside search, or you want to attach an API key for higher rate limits than the anonymous default.
## 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] manages them: openclaw mcp set saves a definition, openclaw mcp list and openclaw mcp show print what's saved, and openclaw mcp status --verbose reports state without starting any servers.
Three details trip people up. First, the transport field is required for Streamable HTTP servers, because OpenClaw defaults to SSE when it's omitted. Second, openclaw mcp set only writes config; running agent sessions won't pick up new tools until you start a new session. Third, OAuth-protected servers authenticate through openclaw mcp login, which stores tokens in a shared SQLite store so parallel OpenClaw processes don't fight over refresh tokens.
openclaw mcp set parallel-search '{"url":"https://search.parallel.ai/mcp","transport":"streamable-http"}'
# verify, then start a new session to load the tools
openclaw mcp show parallel-search
openclaw mcp status --verboseSecurity note
OpenClaw reads inbound messages from your channels, which means every MCP tool you install is reachable by whatever lands in those channels, including prompt injection attempts. Keep per-server approval on prompt rather than approve for anything that can write or spend, and install only servers you trust.
## The best MCP servers for OpenClaw
### 1. Parallel Search MCP
The Parallel Search MCP[Parallel Search MCP] is a hosted server at https://search.parallel.ai/mcp that exposes two tools: web_search, which returns ranked results with token-dense excerpts an agent can usually answer from directly, and web_fetch, which pulls token-efficient markdown from specific URLs, including PDFs and JavaScript-heavy pages. It's free to use with no API key or account, and it runs on our own web-scale index rather than reselling another engine's results.
On the SimpleQA benchmark, Parallel posted 91% accuracy at $8 per 1,000 requests, the strongest accuracy-to-cost ratio among the search APIs we tested (see the published benchmarks[published benchmarks] and methodology). It's also the engine behind OpenClaw's default search, so installing the MCP mainly buys you explicit fetching and, with an API key from platform.parallel.ai passed as a Bearer header, higher rate limits. Our notes on getting maximum search accuracy out of OpenClaw[getting maximum search accuracy out of OpenClaw] 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. The hosted MCP runs Search in its low-latency basic mode and caps excerpts at roughly 25,000 characters per call, so hour-long research jobs belong on the Task MCP below.
### 2. Parallel Task MCP
Search answers questions; sometimes you want OpenClaw to delegate a whole research job. The Task MCP[Task MCP] exposes Parallel's web agents: asynchronous subagents that plan sub-queries, fan out across sources, and come back with a structured, cited result. That covers single deep-research reports and batch enrichment jobs, like researching every company on a list, without burning your main agent's context on dozens of round trips.
It needs an API key, but every Parallel account gets a recurring $5 free monthly allowance[$5 free monthly allowance] plus a signup credit, and pricing is per request rather than per token, so a scheduled research brief costs the same whether the answer took ten pages or a hundred.
**Best for:** scheduled research briefs, competitive digests, and any job where OpenClaw should hand off to a subagent and report back.
**Tradeoffs:** results take minutes, not seconds, and past the free allowance it's a paid API.
### 3. GitHub MCP
GitHub's official hosted MCP server[official hosted 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] 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.
## The picks side by side
| Server | What it gives OpenClaw | Auth | Cost |
|---|---|---|---|
| Parallel Search MCP | Web search + page fetching | None required | Free; API key optional for higher limits |
| Parallel Task MCP | Deep research subagents, batch enrichment | API key | $5/month free allowance, then per request |
| GitHub MCP | Repos, issues, PRs, CI from chat | OAuth | Free |
| Playwright MCP | Real browser automation | Local install | Free |
| Notion MCP | Notes, docs, and databases | OAuth | Free with a Notion plan |
## 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] if you'd rather skip MCP entirely.
A reasonable rule: 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?** No. OpenClaw's default web search is free and powered by Parallel. Add the Search MCP or an API key when you want explicit page fetching or higher rate limits.
**Are these servers free?** Parallel Search, GitHub, and Playwright cost nothing. The Task MCP is covered by Parallel's $5 recurring monthly free credit until you outgrow it. Notion is free with your existing plan.
**Why don't new tools show up after openclaw mcp set?** The command only writes config. Start a new agent session (or restart the current one) and the tools will register. Verify with openclaw mcp status --verbose.
## Start with search
The one-line openclaw mcp set command above is the whole setup: no account, no key, no bill. If OpenClaw becomes the way you work, create a free Parallel account[create a free Parallel account] for higher rate limits and the $5 monthly credit that covers the Task MCP's research subagents. Then judge it the only way that counts: on your own questions.
By Parallel
August 17, 2026