Signal Diff CLI for coding agents

I already know whether a deploy of this blog broke titles, descriptions, or routes. After the Static Web Apps job finishes, Signal Diff crawls the sitemap and compares the live site with a baseline. I wrote that loop up in GitOps with Azure DevOps for .NET apps . The Action is the unattended gate. It sits quiet during the hour I am in Cursor, halfway through a layout change, wanting the same answer before I push.

That is the job of the signaldiff CLI. Coding agents already have a terminal. A terminal is enough to crawl a sitemap and read the report. There is no browser session to open, and no MCP server to configure.

What Signal Diff is doing

Signal Diff fetches the URLs in a sitemap and looks for on-page SEO and crawl problems: titles, meta descriptions, redirects, slow responses, and the rest of the check list. A later run can diff against a baseline, often the previous deploy, so you see what changed.

The site and the GitHub Action came first. I still use those. The CLI is the same product when the caller is a shell, which is where Cursor, Copilot, and Claude Code already are. Signal Diff also has a customer-hosted crawler for teams who want the crawl to stay on their network. This post stays on the coding-agent path.

Why I added a CLI

I wanted a crawl with no account. If I am showing an agent a site, or trying the tool on a sitemap I do not monitor yet, sign-in can wait. signaldiff --sitemap writes an HTML report on this machine. No API key.

I wanted the dashboard questions in the terminal once a key does exist: which sites, what the last run found, what changed since the baseline. The subcommands sites, runs, diff, findings, and scan call the Agent API . The agent stays in the repo. I stay out of a second tab.

An MCP host would have been another install, then another client to teach. The tools I already use can run a command and read a file. One executable on PATH is the setup I wanted. The command reference is at signaldiff.dev/docs/cli .

A few behaviours matter once an agent is the one typing:

• signaldiff with no arguments prints help and exits. It does not start a crawl.
• --json belongs on cloud subcommands. Success JSON goes to stdout. A failure is one JSON object on stderr, with a status code and, on a rate limit, how many seconds to wait. Progress logs stay on stderr, so a pipe keeps the result.
• A local --sitemap crawl refuses --json and still writes HTML.

Install, then crawl this blog

The installers are self-contained. You do not need the .NET SDK, and you do not need a clone of the repo. On Windows:

irm 'https://signaldiff.dev/install/cli.ps1' -OutFile "$env:TEMP\signaldiff-install-cli.ps1"
& "$env:TEMP\signaldiff-install-cli.ps1"

On Linux or macOS:

curl -fsSL 'https://signaldiff.dev/install/cli.sh' | bash

Open a new terminal so PATH picks up the binary (~/.local/bin on Linux and macOS). Then crawl a real sitemap, with a page cap so the first run stays small:

signaldiff --sitemap https://www.funkysi1701.com/sitemap.xml --max-pages 20 --output seo-report.html --quiet

--output creates missing folders before the crawl starts. --quiet keeps the summary short. The report path is on stdout. I point this at my own sitemap because that is the site I already monitor. Swap in yours.

signaldiff update replaces the install with the latest published build and checks the SHA256. signaldiff update --check tells you whether one is waiting, and leaves the current binary alone.

What I tell the agent

The API key stays in the environment. It stays out of the chat. Here is the prompt I give Cursor, Copilot, or Claude Code:

Install the Signal Diff CLI from https://signaldiff.dev/docs/cli . Then, in a new terminal, run signaldiff --sitemap https://www.funkysi1701.com/sitemap.xml --max-pages 20 --output seo-report.html --quiet. Read seo-report.html and list the SEO and crawl issues worth fixing, with the URL for each and a concrete change. Use cloud commands only if SIGNALDIFF_API_KEY is already set. Leave the key out of the chat.

That is the loop. The agent installs if it must, runs a capped crawl, and works from the HTML. I still read the suggestions. A title that is two characters over the limit is a real fix. A rewrite of a post I care about is a suggestion until I agree with it.

This sits next to the split I described in how I use AI on side projects . ChatGPT when the question needs no repo. Cursor when the answer is in the files. The CLI when the question is about the live sitemap.

When a key is already set

Cloud commands need a paid key from the Signal Diff dashboard (Developers → API keys). Set it in the environment. Prefer that to --api-key, which shows up in the process list.

$env:SIGNALDIFF_API_KEY = "sck_…"
$env:SIGNALDIFF_API_BASE_URL = "https://signaldiff.dev/api"

The base URL includes /api. That is the easy one to get wrong. The GitHub Action uses a different variable, and the site origin with no /api suffix. Same secret value. Different names, different URL shape.

These are the commands I want an agent to run once that is set:

runs summary is the one I point agents at. It is a capped brief: headline counts, top regressions, top improvements, and fix hints. It leaves out the full page payload. diff get is there when the only question is what changed against the baseline.

Starting a cloud scan is a separate decision. signaldiff scan start queues work on the service, and signaldiff scan wait <scanId> --json polls until it finishes. I leave that out of the default prompt. A local HTML report answers “what is wrong on this sitemap right now”. A cloud scan spends quota and shows up in the dashboard, so I ask for it on purpose.

The Action still runs after deploy

The pipeline check stays. On this blog the Action still crawls after a Static Web Apps deploy, with the API key in GitHub secrets and a pinned action version. The wiring is in the GitOps post . The Action answers “did this deploy stay healthy?” when I am not looking. The CLI answers the same question while an agent and I are still in the change.

Try it on a sitemap you care about

Install the CLI, cap a crawl with --max-pages, and have the agent read the HTML before it edits templates. If you already have a key, ask for runs summary on the latest run and see whether the brief is enough to act on.

I am curious how you give an agent this kind of check. A CLI it can run, an MCP server, or a dashboard you still open by hand? Tell me in the comments. Command reference, updates, and checksums are on the CLI docs .