Agents, MCP & skills

webx is built for agent harnesses first. Two integration paths — use either or both.

Path 1: MCP server (17 tools)

webx mcp            # stdio transport — point your client at the binary

On a running server (webx serve / webxd), MCP is also exposed over streamable HTTP at POST /mcp.

Client setup

• Claude Code

  # stdio
  claude mcp add webx -- webx mcp
  
  # or streamable HTTP against a running server
  claude mcp add --transport http webx http://localhost:8080/mcp

• Cursor / Windsurf / other MCP clients

  Add to the client’s MCP config (usually ~/.cursor/mcp.json or project .mcp.json):

  {
    "mcpServers": {
      "webx": { "command": "webx", "args": ["mcp"] }
    }
  }

• Devin

  webx mcp   # register as an MCP server in Devin's MCP settings

Tools exposed

search · scrape · ask · query · map · doctor · crawl · extract · diff · research · verify · watch · similar · wayback · llms · answer · batch

Path 2: generated skill

webx skill > .devin/skills/webx/SKILL.md     # or .claude/skills/, .agents/skills/

webx skill emits a SKILL.md that teaches the harness when to reach for webx (search/scrape/index/research) and how to call it — no MCP transport needed, works anywhere the CLI does.

Agent-oriented output features

Feature	Why agents care
--fit <query>	BM25-filtered markdown — drops nav/footer/boilerplate before it burns context
--chunks	Heading-path segments with est_tokens — RAG-ready, budgetable
--a11y + @eN refs	Deterministic element targeting for click:@e3/fill:@e2=x with occlusion checks
--max-tokens	Hard output budget
--agent-ready	Probes llms.txt/AI-bot robots/WebMCP/api-catalog → agent_ready{} report
Prompt-injection warnings	Content-trust flags on scraped output
Web Bot Auth	webx botkey signs requests cryptographically (RFC 9421) — identity instead of stealth

Point agents at a server instead

Any command accepts --api <base-url> to run server-side — agents on thin clients (sandboxes, lambdas) get full capability through one HTTP endpoint:

webx --api http://webx.internal:8080 scrape <url>