Zyte has shipped an MCP server for the thing Scrapy users actually struggle with, which is a crawl that has been running for hours and started doing something odd. In Announcing Scrapy MCP, John Rooney describes a server built on the Remote Control extension in Scrapy 2.19 that “lets the agent discover jobs, check their status and run Python against the live crawler, all without restarting it.” It is the successor to Scrapy’s telnet console, which Rooney says he only got into after talking to Scrapy’s creator about it, redone for an agent. Four tools: list_jobs reads connection files to find running crawls, status reports spider, project, version, process, and uptime, inspection_reference tells the agent what is worth looking at inside a crawler, and execute runs an async Python snippet inside the live process and returns the output.
There are no figures, only requirements and one caveat, and the caveat is the sentence worth keeping. The agent and the crawl must run on the same host under the same user, which keeps the blast radius to one machine. But: “Read-only exploration is an approach, however, not an enforced mode: execute() can also change or stop the running crawl.” The intended use is diagnosis, “an agent can try candidate selectors and compare results before modifying the spider,” and “Scrapy MCP lets an agent inspect what is happening while the problem is present.” The same tool that lets it look lets it act, and the only thing separating the two is the prompt. Installation is one line through the uv package manager, registering the server with Claude Code or a per-project config.
The catalog has to draw a distinction the post does not need to. Scrapy MCP is a local server for the open-source Scrapy framework, not an interface to Zyte’s hosted APIs, so it does not light the MCP server dimension on the Zyte provider page, which lists 3 API pages. The hosted surfaces an agent would reach are the Scrapy Cloud API, where production crawls run, and the Zyte Extract API. The record lights spec presence, error semantics, OpenAPI examples, and agent skills, and agentic access is unlit, so the catalog has not yet mapped which of Zyte’s own operations act.
The Kin Score is 57.3, strong band, carried by developer ergonomics at 76.2, discoverability at 73.2, and operational transparency at 65.8, with access clarity at 61.8 and contract governance at 4.5. The Agent Readiness score is 34.9, agent-ready. Dry-run mode and reversibility are unlit, which is the post’s own caveat read back against the company’s API. Scrapy MCP gives an agent execute with no enforced read-only mode, and Zyte’s hosted contract does not describe a way to rehearse a change or undo one either. The announcement is honest that the guardrail is the approach, not the tool. The record says the same about the platform.