MCP, explained by the 23 tools a publishing platform gives its agents
Latest Model, AI-assisted writing ·
The Model Context Protocol is an open standard for connecting AI applications to external systems [2]. The system publishes a list of tools, each with a name, a description and typed inputs; the model picks a tool and fills in the inputs; the system runs it and returns the result [1]. That is the whole protocol from the model's side. What it looks like in practice depends on which tools a system chooses to publish, so here is ours.
The platform behind the list
The platform publishes 31 small content sites for one owner. It writes, checks and publishes articles; its agents, Claude Code sessions and a custom harness, use MCP tools to read a site's state and act on it. The list below was taken from the platform's code on 2026-11-06 [1].
The 23 tools, in four groups
- 6 read a site's state: list_sites, site_activity, site_floor, site_scorecard, site_topics and article_facts [1].
- 8 act on articles: get_article, queue_draft, archive_draft, restore_draft, unpublish_article, recheck_sources, put_material_page and article_material_page [1].
- 5 change a site's settings: set_author, set_clock, set_look, set_persona and set_provenance [1].
- 4 handle jobs, questions and suggestions: get_job, add_question, list_questions and suggest_persona [1].
An agent never clicks an admin page: it calls site_scorecard to see how a site is doing, recheck_sources to re-read an article's citations, and put_material_page to publish a page of the site's own data [1].
Count your own system's tools by what they do, and the shape shows itself:
- Tools that only read: 6 tools
- Tools that act on content: 8 tools
- Tools that change settings: 5 tools
- Tools for jobs and questions: 4 tools
Works out: Tools in all; Share that only read; Share that change something
What one call looks like
An agent asked "why did this article lose its sources?" calls get_article with the site and the article id, reads the source list and the citations, then calls recheck_sources, which re-reads every source and re-runs the citation check. The agent reports what changed. It needed no database access and no login: only the tool list and a platform key [1].
What is deliberately missing
No tool lets an agent approve its own draft. A draft posted with queue_draft goes live only if the quality gate passes it and the site's rule allows it (publish automatically, or wait for the owner); otherwise it is held [1]. That absence is the safety design: what is left out of the list, plus the gate behind every write.
What MCP does not give you
- Judgement. A tool runs what it is asked; deciding when to archive or unpublish stays a rule or a person [1].
- Safety by itself. Here the safety comes from what is left out (no self-approval) and from the gate behind every write, not from the protocol [1].
What this does not show
One platform's tool list on one date. Tool names may change, and the grouping is ours [1].
A notebook entry for Latest Model, written with AI from the platform's own code listing and checked by hand. The tool counts are the platform's; the calculator computes the shares from any counts you enter.