fetch still has to work out, on its own, whether fetch is even the right call for this page, how to start a real project instead of a one-off snippet, how to script something that runs the same way every time, and how to tell whether a change quietly broke something. The Agent Toolkit is the part of the Zenrows CLI that answers those questions directly, so an agent (or the developer directing one) doesn’t have to reconstruct them from scratch on every project.
Each piece solves one specific part of that problem:
init --all sets up a full workspace: every skill, template, workflow, recipe, and eval, plus the MCP config for your agent clients, printed ready to paste, in one command. plugin install <client> is the narrower path for a single client: it installs the Zenrows skills into .zenrows/skills/ and prints the MCP config to paste into that client; it doesn’t edit the client’s config files. Check what’s already in place at any point with zenrows status.
The Shape of the Problem
fetch, extract, batch, and browser entirely on its own, no agent client, no MCP:
- Connect an MCP-native agent client, Claude Code, Cursor, and similar, so the model can discover and call tools directly instead of a human typing CLI commands. MCP sets this up. Only relevant if you’re wiring in that kind of client, an agent that already runs shell commands (like one driving the CLI directly) doesn’t need it at all.
- Decide which primitive fits the task, without guessing. Skills are short, agent-readable playbooks, a master decision tree plus one playbook per primitive, so an agent picks
fetch,extract,batch, or a browser session correctly, without extra prompting from you. - Start a real project instead of an empty file. Templates scaffold a minimal, working project for a specific job, like calling Fetch from Node.js or building a Batch job spec.
- Automate a multi-step process. Workflows are playbooks: the shipped one, for keeping competitor data fresh, is read with
zenrows workflow explain, and you script its steps yourself with the primitives if you want to repeat them. - Verify that a single command actually works, right now, against a known-good target. Recipes are small, runnable examples that double as a smoke test before you build anything on top of them.
- Benchmark that it keeps working. Evals run the same check reproducibly and write a scored report, so a regression shows up as a failed eval instead of a silent surprise in production.
The Pieces
Skills, templates, workflows, recipes, and evals share
list, install, explain, update, and remove. Each adds its own verbs: skill validate and generate, template create, recipe run, and eval run and report. The shipped workflow is read with explain. MCP has its own commands, zenrows mcp and zenrows plugin.
No account or API key is needed to start. The CLI’s first cloud command, like the local MCP server’s first tool call, creates a Free account with 5,000 credits for its first billing period when no key is set, and you can claim it later, keeping its key and usage. An unclaimed account is not renewed after that first period. See Agent signup and unclaimed accounts. An agent can also pay per request with no account at all, buying and calling Fetch and Extract over x402, MPP or a card at
https://agents.zenrows.com/llms.txt. See Agentic Payments.Next Steps
MCP
Connect Claude Code, Cursor, and other clients.
Skills
See the full catalog of agent playbooks.
Templates
Scaffold a working project.
Workflows
Read a multi-step playbook with
workflow explain.Recipes
Run one example end to end.
Evals
Benchmark a command with a written report.