Appearance

Use EdgeComet MCP with coding agents ​

Coding agents such as Claude Code, Codex, GitHub Copilot and Cursor can read your production SEO data through the EdgeComet MCP server while they work in your repository. The agent checks what Googlebot and AI bots received before it changes a template, and checks the live result after the deploy.

Without it, an agent changing SEO code works from the source alone. It cannot see which pages bots fetch, which tags they received, or what the last release broke.

Before you start ​

Connect your agent to the server: Claude Code, Codex, GitHub Copilot or Cursor. Then ask it to list your websites to confirm the connection works.

Tell the agent to use EdgeComet ​

Add a short block to the instructions file your agent reads, such as AGENTS.md or CLAUDE.md at the root of the repository. Replace the domain and website ID with yours: the agent finds the ID with list_websites.

markdown
## SEO data (EdgeComet)

Production SEO data for example.com is available over the `edgecomet` MCP server (website ID 42).

- Before you change templates, metadata, internal links, redirects or robots rules, check production data in EdgeComet. Do not guess what search engines or AI bots receive.
- For site-level numbers, call `list_reports`, then `get_report` with its provider keys. Quote those numbers instead of recomputing them.
- Before you write a `query`, call `get_schema` and use its field names exactly.
- Quote every number with the window and crawl the result reports.
- After a deploy, check the live page with `read_page`.
- Draft Edge SEO rules only when asked. A person deploys them in the EdgeComet dashboard.

The block turns the server instructions into project rules: orient first, use the schema, verify live.

Fix a template from production data ​

  1. Ask the agent for the problem and where it lives, for example: "Which indexable pages render without a canonical tag, grouped by folder?"
  2. The agent queries the per-page data and finds the folders and page types involved. It reads one page of each type with read_page to see the served markup.
  3. It finds the template in your repository that renders those pages and fixes it.
  4. After the deploy, it reads the same pages with read_page to confirm the tag is in what bots receive.

Over the next days, the page change log records the new value as bots fetch the pages. Ask "did the canonical change reach the pages Googlebot fetched since the deploy?" to check coverage.

Check a release after it ships ​

EdgeComet detects a release when the same change appears on many pages of one template. After a deploy, ask the agent:

What did the last release change, and did any pages lose indexability?

list_releases returns each release with its class (loss, edit or gain), its page types and example pages with old and new values. The agent can match a lost title or canonical to the commit that removed it. See Site releases.

Bridge a fix with Edge SEO ​

When the code fix waits for a release cycle, the agent can draft the same change as an Edge SEO rule that serves the corrected tags to bots now:

  1. The agent maps the page with render_page_map and reads the target element with read_page and selectors.
  2. It previews the change with preview_edge_seo_changes and saves a draft.
  3. It gives you the editUrl or reviewUrl. You review and deploy the draft in the dashboard.

When the code fix ships, remove the rule. See Edge SEO drafting.

Let read-only tools run without prompts ​

Sixteen EdgeComet tools carry MCP's read-only hint. Each client handles approvals in its own settings, as their docs describe them:

  • Claude Code: allow tools with permission rules in .claude/settings.json, such as mcp__edgecomet__query for one tool or mcp__edgecomet__get_* for a group. A connector added on claude.ai shows up as claude.ai edgecomet, and its tools are named mcp__claude_ai_edgecomet__<tool>.
  • Codex: set default_tools_approval_mode = "writes" on the server, which prompts only for tools that are not marked read-only.
  • GitHub Copilot in VS Code: VS Code's MCP guide says read-only tools run without confirmation, and your permission mode can change that. In Copilot CLI, every MCP tool asks unless you pass --allow-tool='edgecomet'.
  • Cursor: asks before MCP tools by default. In Auto-review mode, tools on your allowlist run at once.

The Codex setting in ~/.codex/config.toml:

toml
[mcp_servers.edgecomet]
url = "https://cloud.edgecomet.com/mcp"
default_tools_approval_mode = "writes"

The list of read-only tools is in the tool reference.

Keep the agent's context small ​

  • Ask for computed numbers first. One get_report call returns up to 20 dashboard numbers. Rebuilding them from rows costs many calls.
  • Aggregate before you list. A query with group_by returns up to 200 groups. A row listing returns 50 rows a page.
  • Export large results. export writes up to 1,000,000 rows to a CSV link instead of the conversation.
  • Map pages, do not dump them. render_page_map outlines a page's structure. Reading a whole page as HTML to find one selector fills the context.

Next ​

Related Articles