What the system can doAI models and agents
One family of actions
AI models and AI agents
Ask the AI model you choose, connect AI apps to the Directory, and hand tasks to AI agents that report back.
Any AI model that can use tools (Claude, ChatGPT, Grok, Gemini) can run the Directory’s actions. AI agents are separate programs, each running an AI model with its own Directory and Ledger.
Live, read from this site just now.
The groups in this family, counted just now: — actions
This family among all — actions in the Directory, one square each
Each action is one row in the Directory, sorted into this family by the start of its name.
What it can do
What the actions in this family do, group by group.
Each card is a group of actions that do similar work. The number on each card is how many actions are in that group right now.
Switch to “For your developer” to see the real name of every action.
The real name of every action, read from the Directory just now. Names that contain a supplier’s name are left out of the lists, and still counted.
Ask a model
—Send a question or a task to the model you choose, including open models running on Cloudflare; translate text; turn text into numbers for search by meaning; ask several models the same question.
Models in a signed-in browser
—Ask ChatGPT, Claude, Grok or Gemini through their own websites, in a browser signed in to your own subscription, and keep the full answer; or ask three at once.
Choose models and set limits
—List the models available, choose how hard a model thinks, edit an assistant’s written instructions, and cap how many actions, how much memory and how much spend one turn of a model may use.
Connect AI apps (MCP)
—MCP (Model Context Protocol) is the standard way AI apps connect to outside tools. Connect Claude Code, Codex and other AI apps to the Directory, and connect outside MCP servers so their tools become rows too.
AI agents on Cloudflare
—Separate programs on Cloudflare, each running an AI model with its own Directory, Ledger and spreadsheet: give one a task, read its answer and its record, change its settings, or make a copy of it.
Coding assistants on the Mac
—Start a coding assistant (Claude Code, Codex and others) on the Mac with a task, send a message into a running session, and read what each one did.
Agents that keep working
—Start an agent that keeps working on a goal until it is done, pass messages between agents, let two agents argue a plan until they agree, and save the lessons an agent learns.
Uses
What a business uses these actions for.
Examples of the work these actions do. They describe uses, not results.
Ask two models the same question and compare their answers in the Ledger.
Change the model that answers your customers without rebuilding anything.
Give an AI agent a task in a spreadsheet row and read its answer in the next column.
How it is run
How any of these actions is run.
The same way as every other action in the Directory: one request, the same permissions, one record.
You don’t have to pick an action yourself. You text what you want; the AI assistant picks the action and runs it within the permissions you set. The result comes back to you, and the request, the result and who asked are written to the Ledger, the record of everything the system has done.
Every action is one row in the Directory, and each row can be run — ways: a text, a link, a spreadsheet cell, an iPhone Shortcut, a command, or an AI model’s request. Every way goes through dispatch: one web address that looks the action up, checks the permissions, runs it on the right computer, sends back the result and writes a Ledger entry. The same action answers every way, under the same limits.
Dispatch takes an HTTP POST (a standard web request) of {"key": "<action name>", "args": {…}}. The same key works over the API (the web address your software calls), MCP (Model Context Protocol, the standard way AI apps connect to outside tools), the command line (CLI), a webhook (an address another program posts to), a spreadsheet cell or a Shortcut. A token (a key or a link) can be limited to exactly the actions you allow; every call, and every refusal, is a Ledger entry.
// One action from this family, as dispatch receives it
{ "key": "…", "args": { … } }Ask for this.
The text below opens ready to send. Add a line about your business, and I’ll reply with what this family of actions would do for it.
Every family
Everything else the system can do.
Every family of actions, with how many actions it has right now.