Connect an agent

A connected agent reads your documents, writes versions and comments, and asks you to approve its plans, all as you, under its own name: Claude Code (you@example.org). readthis.md supplies no model; connect the one you use.

Connect

Each way below opens your browser: sign in to readthis.md and press Allow. The lines are for readthis.md; on your own server, use its address in place of https://readthis.md.

Claude Code, with the readthis.md plugin, which also puts plan mode's plans on readthis.md. On your own server, add --config readthis_url=https://your.server before the &&.

claude plugin install readthis --marketplace LiamBindle/readthis.md && claude mcp login plugin:readthis:readthis

Codex, with the skill that has it put every plan to you:

codex mcp add readthis --url https://readthis.md/mcp && mkdir -p ~/.agents/skills/readthis && curl -fsSL https://readthis.md/skill.md -o ~/.agents/skills/readthis/SKILL.md

Claude, on the web or the desktop app: Customize → Connectors → Add custom connector, and paste https://readthis.md/mcp.

ChatGPT: chatgpt.com/plugins → + → Add custom MCP server. Paste https://readthis.md/mcp and choose OAuth.

Cursor: put this in ~/.cursor/mcp.json, then add the skill:

{"mcpServers": {"readthis": {"url": "https://readthis.md/mcp"}}}

mkdir -p ~/.agents/skills/readthis && curl -fsSL https://readthis.md/skill.md -o ~/.agents/skills/readthis/SKILL.md

Gemini CLI: run this, then /mcp auth readthis in Gemini CLI, then add the skill as for Cursor.

gemini mcp add --transport http readthis https://readthis.md/mcp

Anything else that speaks MCP takes the address https://readthis.md/mcp. If it reads skills, add the skill as for Cursor; if it can't sign in, give it an API key.

Plans

Before a connected agent starts work, it puts its plan up as a document and asks Go ahead with this version? Comment on what is wrong and press Request changes: it replies in each thread, revises, and you compare the versions. Approve lets it start.

Revising a document

The same loop works on any document, or only some of its pages. Several agents can work at once on different pages: their saves stack, and each asks about its own. In Claude Code, /readthis:revise runs the loop.

Skills

Tag a document skill and it becomes one of your skills. When a session starts, a connected agent sees the name and description of each of your own skills, and reads one when it’s relevant. Its description says when to use it. One somebody shares with you is not listed, since its words would reach your agent as if they were yours.

Pointing an agent at something

Copy a link to a document or a page from its ⋯ menu, to a draft or a version from its row in Drafts or Versions, to a section from the # beside its heading, or to a comment from its ⋯, and give it to your agent. It opens what the link points at.

API keys

A script, or an agent that cannot sign in, uses a key you create under Agents in your settings, behind your address. It is sent as Authorization: Bearer <key>, and acts as you until you remove it. Every agent you connected is listed there too, and removing one disconnects it.

What an agent can never do

An agent can never answer a request meant for you, bring anyone new in, let anyone edit, remove access, move a document to the trash, delete it, download a copy, make a key, or change your settings. Only you can do those, in the browser.