Best Web Data Plugins for Codex and Cursor in 2026

Coding agents now have two places to get web access from: MCP servers you add by hand, and plugins that bundle skills, an MCP server, or both into one install. Codex and ChatGPT share a single plugin directory; Cursor has its own marketplace. Firecrawl published a run of "best plugins" roundups in the first week of October 2026 and put its own plugin first in each. This post asks a narrower, checkable question: for the four web-data vendors we compare elsewhere, what can you actually install today, and what comes with it?

We installed each one, so the answers below come from a run, not from a README. Where we could not test something, we say so. And one of the four is us.

What we tested, and what we did not

All installs ran on October 13, 2026 with codex-cli 0.156.1, each into a fresh, empty CODEX_HOME so nothing carried over between vendors. We used codex plugin marketplace add, codex plugin add, then looked at what landed on disk and what codex mcp list showed.

We did not run a model session, so we cannot tell you how well any plugin's skills perform. We did not test Cursor at all: it is not installed on the machine we used, so Cursor claims below come from Cursor's documentation and each vendor's repo. And we could not check directory listings from outside the apps.

What to evaluate

• Where it lives. A directory listing (one click, reviewed) or a GitHub repo (two commands, you read it first).
• What it bundles. Skills only, an MCP server, or both. Skills teach the agent a workflow; an MCP server gives it live tools.
• Auth. OAuth sign-in, an API key in an environment variable, or none. This decides what a plugin is able to bundle.
• License. MIT, a vendor EULA, or unstated. It governs whether you can fork and review the skills for your team.
• Client fit. Some packages are built for one client and only load elsewhere by accident of format.

The contenders

Two of those marketplace adds worked through a back door worth knowing about. Codex reads a Claude-style .claude-plugin/marketplace.json as a legacy location, so Bright Data's repo (published as a Claude Code plugin) and Firecrawl's CLI repo both installed in Codex without a Codex-specific manifest. Firecrawl's separate firecrawl-codex-plugin repo, by contrast, could not be added as a marketplace directly: Codex reported no supported manifest at its root, and its README says to wire up a marketplace file yourself.

Firecrawl: the one with directory listings

Firecrawl ships a plugin per client: separate repos for Claude Code, Codex, Cursor, and others. The Cursor plugin bundles a hosted MCP server at https://mcp.firecrawl.dev/v2/mcp-oauth that opens a browser sign-in on first use (no key to paste), plus 13 skills and an install rule. Firecrawl's own roundup says to install it in Cursor with /add-plugin firecrawl and links its Cursor Marketplace page, and its October 6 post announces an official ChatGPT plugin that it says also works in Codex. Those listings are the real differentiator, and we have not verified them.

What we did verify is the GitHub route: the firecrawl/cli repo installs in Codex as a 13-skill plugin; Firecrawl's plugin READMEs describe those skills as working through the Firecrawl CLI. Its dedicated Codex repo describes itself as skills-only and says, as of its last update in June, that it was waiting on a directory listing, so read Firecrawl's October posts and that README together. Choose it if you want the one-click path and the broadest client coverage, and you are happy to sign in with OAuth.

Bright Data: a Claude Code plugin that happens to load in Codex

Bright Data's brightdata/skills repo is titled a Claude Code plugin, MIT-licensed, and updated October 12. It installed in Codex as 21 skills: scraping, search, data feeds, scraper building, proxy, and SDK best-practices guides. It bundles no MCP server. In Codex that means you add Bright Data's MCP separately if you want the tool surface, and we did not find a Cursor-specific package. Choose it if you already work with Bright Data's infrastructure and want its skills in the agent.

Zyte: the only bundled MCP, because it uses OAuth

Zyte publishes dedicated packages for Claude Code, Codex, and GitHub Copilot CLI, plus a generic agent-plugin repo (zytedata/skills) for other agents. The generic repo could not be added as a Codex marketplace; the dedicated zytedata/codex-skills repo could. It installed with five skills and registered Zyte's hosted MCP server, which listed as "Not logged in", matching Zyte's OAuth sign-in. Zyte's skills generate a working Scrapy project from a plain-English prompt, which is a different job from a tool that returns data. Cursor appears in Zyte's marketing copy and its generic repo's README, but we found no dedicated Cursor package. Choose it if what you want out of the agent is a maintainable scraper codebase, and you are fine with the Zyte EULA.

Crawlora: the broadest skill set, skills-only in Codex

Crawlora's plugin is the repo itself: one crawlora plugin that installs all 165 skills, which call Crawlora's REST API with your CRAWLORA_API_KEY. To get the agent-native MCP tools in Codex, add the server separately:

codex plugin marketplace add Crawlora-org/crawlora-skills
codex plugin add crawlora@crawlora
export CRAWLORA_API_KEY=sk_your_key_here
codex mcp add crawlora --url https://mcp.crawlora.net/mcp --bearer-token-env-var CRAWLORA_API_KEY

For Cursor the repo is also a plugin: skills load from skills/, and a root mcp.json registers the hosted MCP server with "Authorization": "Bearer ${env:CRAWLORA_API_KEY}", the interpolation syntax Cursor documents. We have not installed it in Cursor. It is not listed in the Cursor Marketplace or the Codex directory, and we have not submitted it to either. Choose it when the work is structured records from named platforms (prices, listings, filings, profiles) across many sources and you would rather install from a repo you can read.

Why only OAuth servers can ride inside a Codex plugin

This is the part we did not expect to find. We pointed a plugin's bundled mcp.json at a local mock server and logged what Codex sent. Custom headers arrived. But an Authorization header never did, and a value like ${CRAWLORA_API_KEY} was sent as that literal text, not expanded from the environment. Keys such as bearer_token_env_var caused Codex to drop the whole server. So a server that authenticates with a static API key has no way to ship a working credential path inside a Codex plugin today, and the vendors who bundle MCP there are the ones using OAuth. That is why Zyte's plugin registered a server and the other three did not.

The portable agent-plugin schema agrees: it defines only type, url, and headers for a remote server and no environment interpolation. Cursor's documentation takes the opposite approach and does resolve ${env:NAME} in url and headers. Expect this to differ by client and to change; these results are from one Codex version on one day.

Directory or GitHub: how to choose

• One-click for yourself or a small team. A directory-listed plugin (Firecrawl, per its own announcements) is the shortest path.
• A team that wants to read what it installs. A GitHub plugin you can pin and review. All four vendors qualify; Crawlora and Bright Data are MIT, Zyte is under its own EULA.
• A scraper codebase out of the agent. Zyte's plugin.
• Many named platforms behind one key. Crawlora's skills, with the MCP server added separately in Codex.
• A hard, named site at volume. Bright Data's skills plus its infrastructure.

Install Crawlora in Codex or Cursor

165 skills and a hosted MCP server, one API key, billed only on success. 2,000 free credits a month, no card.

Set up Crawlora MCPBrowse the API catalog

Sources

Related reading

• Best MCP Servers for OpenAI Codex in 2026 — the same four vendors as MCP servers, including Zyte's Codex add-on.
• Best AI Agent Skills Marketplaces in 2026 — where skills are listed and discovered.
• Best Web Search MCP Servers in 2026 — live-tested search endpoints.
• Crawlora vs Firecrawl · Crawlora vs Bright Data · Crawlora vs Zyte.
• Give your AI agent live web data with MCP — how MCP works and how to connect in three minutes.