Best MCP Servers for Warp in 2026

Warp turned its terminal agent into a standalone CLI this month. One binary now runs inside Ghostty, iTerm2, VS Code, your native OS terminal, or over SSH with no remote install. That's a real expansion of where Warp shows up in a developer's stack, and nobody's published a Warp-specific buyer's guide for web-data MCP servers yet. This one compares what actually connects, with real config, not marketing copy.

Warp's MCP config

Warp reads MCP server definitions from a JSON file at ~/.warp/.mcp.json (global, applies everywhere) or .warp/.mcp.json at your project root (project-scoped, useful for keeping a client's API key out of servers you don't want available in every session). The root key is mcpServers, the same shape Claude Desktop uses, so most existing configs paste in with no changes. Warp also picks up config from a few third-party agent providers directly (~/.claude.json, ~/.codex/config.toml) if you'd rather not maintain a separate file.

You don't have to hand-edit the file, either: Cmd+P / Ctrl+P → "Open MCP Servers", or Settings → Agents → MCP servers, opens a UI that edits the same JSON for you and shows a live connection status per server.

Warp supports two server shapes:

• Command-based (local/CLI): { "command": "...", "args": [...], "env": {...}, "working_directory": "..." }: runs a local process, typically via npx.
• Streamable HTTP/SSE (remote): { "url": "...", "headers": {...} }: points at a hosted endpoint, with a headers field for an Authorization bearer token.

Where Warp differs: provider configs, trust gates and logs

We re-read Warp's MCP page on October 26, 2026, and three behaviors on it matter more for a web-data server than the JSON shape does.

Warp reads other agents' config files. Besides its own file, it picks up the MCP definitions Claude Code and Codex already keep on disk, plus a generic agents location:

If a Firecrawl, Bright Data or Crawlora entry already works in Claude Code, you can enable Auto-spawn servers from third-party agents and skip writing it a second time.

A cloned repo cannot start a server by itself. Warp detects project files but never launches their servers automatically, and you start each one by hand from the MCP servers page. That choice lasts for the session, so after a restart you switch the server on again if you still trust the repo. Edits to the config files also need your approval. For a client project that keeps its own API key in .warp/.mcp.json, this is the behavior you want, and it means a freshly checked-out repo will show the server as stopped, not broken.

Credentials follow three routes, and sharing treats them differently. Warp's docs describe env variables, a browser OAuth sign-in that saves credentials on your device (you sign in again on a new machine and can revoke from the MCP pane), and a bearer token in headers. When you share a server with a team, Warp's docs say sensitive env values are replaced with variables that each teammate fills in. They say nothing comparable about headers, so check what a shared copy contains before sending a hosted server that carries its key in a header. A hosted server with browser sign-in, such as Firecrawl's OAuth URL or Zyte's, could in principle use the OAuth route; we have not run either in Warp.

Debugging is local. Each server on the MCP servers page has a View Logs button, and the same files sit on disk: under ~/.local/state/warp-terminal/mcp on Linux (or $XDG_STATE_HOME), and in a folder named mcp inside Warp's application support directory on macOS. Warp warns that those logs may contain API keys, so strip them before pasting into an issue. If sign-in gets stuck, deleting ~/.mcp-auth clears every stored MCP token on the machine, and Warp's own tip is to try a different model when a server's tools are never called.

What to evaluate, in a terminal

• Process or no process: a command-based server means an npx child process per Warp session (and per SSH host if you use the Agent CLI remotely); a url-based server is one line of JSON and nothing running locally.
• Where the key lives: Warp's project-scoped .warp/.mcp.json lets you keep a client's API key out of your global config, which matters more in a terminal you SSH around with than in an IDE.
• Generic crawl (any URL → markdown) versus structured platform data (named sites → JSON). Terminal agents tend to do both: read a doc page, then pull a number from a known site.
• Tool count versus tool relevance: 60 tools that all need a browser, or 4,232 that map one-to-one onto platform endpoints? Warp's tool picker shows them all; more is not automatically better.
• Billing on failure: a blocked page in a scripted terminal loop can retry many times. Whether each failed attempt is billed changes the cost of a bad afternoon.

The contenders

Firecrawl MCP: the "read this page for me" tool

Firecrawl's MCP server exposes scrape, crawl, map, search, extract and interact tools against any public URL, with no fixed catalog. In Warp it is a command-based entry (npx -y firecrawl-mcp, FIRECRAWL_API_KEY under env), which means an npx process starts with each Warp session that uses it; over SSH with the Agent CLI, that process runs wherever the agent runs. Choose it when the terminal task is "pull that docs page or README into context", which is the single most common web call a coding agent makes.

Bright Data: anti-bot scale, and the earliest Warp guide

Bright Data documented a Warp integration months before the standalone Agent CLI existed, and the same command-based config (npx @brightdata/mcp, API_TOKEN under env) still applies. Older guides describe a free Rapid tier and a PRO_MODE: "true" upgrade; Bright Data removed PRO_MODE in v3.1.0 on October 15, 2026, so the server now exposes all 32 of its tools by default, covering search, scraping, 1,200+ pre-built scrapers and its browser infrastructure. Choose it when the target is a hard, named site and anti-bot reliability at scale is the constraint, and you are fine running a local process to get it.

Zyte: a hosted server now exists, Warp isn't documented

Update, October 5, 2026: Zyte launched a hosted MCP server on September 30, 2026 (https://mcp.zyte.com/v1/mcp), which this post first missed. Zyte's docs cover other clients, not Warp, so it is untested here. In Warp it would be a url entry in .warp/.mcp.json; whether Warp's remote entries can finish Zyte's sign-in is the part we have not checked. Tools and billing: see our Pi write-up.

Crawlora: one JSON line, no process, 497 platform groups

Crawlora's MCP server is the only url-based entry in this comparison: point Warp at https://mcp.crawlora.net/mcp with Authorization: Bearer <key> under headers and there is no npx to install, update or keep alive. In a terminal that is the practical difference: the same .warp/.mcp.json works on your laptop, in VS Code's terminal, and on a remote box over SSH with no remote binary. The 4,232 tools map one-to-one onto documented REST endpoints across search, maps, e-commerce, social, finance, travel, jobs, media, app stores and reviews, return normalized JSON, and are billed only on success, with 2,000 free credits a month and no card. Choose it when the agent needs a number or a record from a named platform (a price, a ranking, a listing, a filing) rather than a page to read.

{
  "mcpServers": {
    "crawlora": {
      "url": "https://mcp.crawlora.net/mcp",
      "headers": { "Authorization": "Bearer YOUR_CRAWLORA_API_KEY" }
    }
  }
}

How to choose, by job

• "Read this page / these docs into context" → Firecrawl MCP.
• "Get me the data off this hard, named site, at volume" → Bright Data, and accept the local process.
• "Give me the price / ranking / listing / filing from a known platform, as JSON, without installing anything" → Crawlora MCP.
• "I want Zyte's Scrapy-generation workflow" → Claude Code or GitHub Copilot for now, not Warp.
• Working over SSH with the Agent CLI → prefer url-based servers; a command-based one runs on the remote host and needs Node there.

Add structured web data to Warp

4,232 tools across 497 platform groups, 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 JetBrains IDEs in 2026: another client with a recent AI-agent update, same three vendors.
• Best MCP Servers for GitHub Copilot in 2026: the same comparison, Copilot's config shape.
• Best MCP Servers for Windsurf in 2026.
• Best Web Scraping MCP Servers & Skills for Claude Code in 2026.
• Crawlora vs Firecrawl · Crawlora vs Bright Data.
• Give your AI agent live web data with MCP.