Let an MCP agent do SEO research, then stop at a CMS draft

Four read-only research tools and a nine-step playbook that ends at an approved draft.

  • MCP connectionNo keyOAuth or a site-scoped token on your instance.
  • ScopeNo keycontent:read for the four research tools.
  • Search keysKeyWithout a Brave key, research returns only the question tree.
4
read-only research tools
9
playbook steps
0
publish tools
1
approval before a draft
MCP client · BlogFactory · agent playbookIllustrative composite

Agent session

  1. 1
    discover_topicssite_id: plantwise · 24 topics, 5 with overlap
  2. 2
    research_keywordquestion tree + SERP brief (Brave key set)
  3. 3
    generate_draftasks first: spends the provider budget
  4. 4
    analyze_postreadability, density, structure, near-duplicates
  5. 5
    update_draftsends expected_updated_at from the latest read
  6. 6
    review_postReviewPacket: preflight, destinations, warnings
  7. 7
    push_to_cms_draftonly after approval of this version

Latest result

Not called. The agent stops here and asks you to approve.

Four research tools are read-only. generate_draft asks first because it spends your provider budget. Publishing is not a tool.

Illustrative composite of an agent session. Step through it with the controls. Sites, topics and results are invented.

The agent loop

  1. Discover topics

    discover_topics returns ranked topics from the saved profile.

  2. Research the keyword

    research_keyword returns the question tree and, with a Brave key, a SERP brief and a plan_brief for drafting.

  3. Generate, analyze, fix

    generate_draft spends your provider budget, so the agent asks first.

  4. Review and wait

    review_post shows blockers.

What it needs
MCP connectionNo keyOAuth or a site-scoped token on your instance.
ScopeNo keycontent:read for the four research tools.
Search keysKeyWithout a Brave key, research returns only the question tree.
What it will not do
  • Publish live, delete content, or administer accounts.
  • Return or accept credentials in a prompt.
  • Treat search results, competitor pages or sitemaps as instructions; they are data.

BlogFactory never publishes live. Agents propose, people approve, and the furthest an agent goes is a CMS draft.

More: what you get, what changes, limits and questions

An MCP client can do the research half of SEO work through four read-only, site-bound tools, then draft, analyze and revise through the existing draft tools. The agent playbook lists the loop, and its last step needs you: an agent can request a CMS draft only after you approve the exact version, and it can never publish.

What you get

Four read-only tools

analyze_post, discover_topics, research_keyword and find_content_gaps need only content:read, stay inside the approved site and change nothing.

A playbook, not a script

Discover, research, generate, analyze, fix, review, then wait for approval. Each step names the tool and what to do with its result.

Authority that stops

The highest agent authority is delivering an approved version as a CMS draft. Publishing, deleting and reading credentials are outside it.

What changes

Before

  • Research steps are manual prompts with no shared procedure.
  • Agents can reach for tools they should not use.
  • Nobody knows where the agent stops.

With BlogFactory

  • Four read-only, site-bound research tools.
  • A nine-step playbook from discovery to review.
  • The agent stops before push_to_cms_draft until you approve the exact version.

OutcomeA reviewed version waiting for your approval.

Limits to know

  • The research tools read fixed public sources: Google autocomplete, your Brave key, public pages and a competitor's public sitemap.
  • Private and internal hosts are refused. No provider credential is ever returned.
  • Discover the live tool catalog with tools/list rather than assuming a fixed list.

Questions

Can an AI agent publish with BlogFactory?

No. The furthest an agent can go is to request a CMS draft for a version you approved. Publishing stays a human action.

Which MCP clients work?

Any client that supports Streamable HTTP MCP with OAuth or a bearer token, including Codex, ChatGPT and Claude. The agent guide has connection commands.

Do the research tools change my site?

No. analyze_post, discover_topics, research_keyword and find_content_gaps are read-only and write nothing.

Where is the playbook?

In the docs, as the agent playbook, with the prompt to paste into a client or keep as a skill.

Draft-only, whether you self-host or use Cloud.

Cloud starts at $5 a month and AI usage is bring-your-own. The open-source core is free to self-host. The authority boundary is identical in both.