Access · MCP
Connect Claude, ChatGPT, Cursor or Claude Code to every action in the Directory
MCP (Model Context Protocol) is the standard way AI apps connect to outside tools. The system runs an MCP server. Connect an AI app to it once, and every action in the Directory appears in that app as a tool it can use, under the same permissions, with every call written to the Ledger.
Connecting needs a key from us.
3 tool calls asked for, 2 ran, 1 refused: 3 Ledger entries
An illustration. The business and the people are made up.
Live, read from this site just now.
How much detail
Developers: choose “For your developer” for the protocol details and the raw requests.
What connecting does
The AI app sees every action as a tool, and asks the system to run it
- 1
You add the address and your key
Once, in the AI app’s settings or with one command.
- 2
The app lists the tools
One tool for each action in the Directory.
- 3
You ask in your own words
The AI picks the tool and fills in its inputs.
- 4
The system checks and runs it
The same permissions as every other way in. Past a limit, it is refused.
- 5
The call is written down
What the AI asked for and what came back, as Ledger entries.
Your AI app can do what the system can do, but only what your key allows, and you can read afterwards exactly what it asked for.
Claude, ChatGPT, Cursor and Claude Code reach the same actions with the same limits. If you switch AI app, the actions, the permissions and the record stay the same.
Streamable HTTP, stateless, JSON answers, at https://miscsubjects.com/mcp. tools/list returns one tool per Directory row; tools/call runs the row through the same runner as every other way in, and each call and each answer is a Ledger row. Protocol versions 2025-06-18, 2025-03-26 and 2024-11-05.
Connect
How to connect each app
Each app needs the same two things: the address https://miscsubjects.com/mcp, and your key.
Claude Code
Claude Code is Anthropic’s coding assistant for the terminal. Add the server with one command.
claude mcp add --transport http miscsubjects https://miscsubjects.com/mcp \ --header "Authorization: Bearer YOUR_KEY"
Cursor
Cursor is a code editor with an AI assistant. Add the server to its MCP settings file, mcp.json.
{
"mcpServers": {
"miscsubjects": {
"url": "https://miscsubjects.com/mcp",
"headers": { "Authorization": "Bearer YOUR_KEY" }
}
}
}Claude and ChatGPT
Both apps add outside tools as connectors in their settings, using the address above. Tell us which app you use when you request access, and we set up how that app signs in with your key.
Any other MCP app
Any app that speaks MCP over HTTP connects with the address and the header Authorization: Bearer YOUR_KEY.
Choose “For your developer” to see the raw requests an app sends.
curl -X POST https://miscsubjects.com/mcp \
-H "authorization: Bearer YOUR_KEY" \
-H "content-type: application/json" \
-H "accept: application/json, text/event-stream" \
-d '{"jsonrpc": "2.0", "id": 1, "method": "tools/list"}'curl -X POST https://miscsubjects.com/mcp \
-H "authorization: Bearer YOUR_KEY" \
-H "content-type: application/json" \
-H "accept: application/json, text/event-stream" \
-d '{"jsonrpc": "2.0", "id": 2, "method": "tools/call",
"params": {"name": "NOW", "arguments": {"args": {}}}}'YOUR_KEYis a placeholder: you get a key from us.
Permissions
An AI app gets no more than its key allows
The AI chooses which tool to call. The system decides whether the call runs.
- Every tool call passes the same permission check as any other request: the key, the action, and the action’s own limits.
- A call past a limit is answered as refused, and the AI is told why.
- A refused call is written to the Ledger like any other.
- Switch an action off and it is off for the AI too, in every app.
What the AI asked for
Refund everyone who booked last month
The action’s limit
One customer per refund, up to $500 without you
What happened
Refused. Nothing ran. The refusal is a Ledger entry.
An illustration. The business and the people are made up.
Without MCP
Other ways an AI model can call an action
A model’s function call, as the model wrote it
Many AI models write a tool call as a small block of JSON. The system accepts it exactly as written, in the shapes used by OpenAI, Workers AI, Anthropic, Gemini and MCP, runs the action, and answers with the result written back in the same model’s shape.
A tag inside the model’s text
A model that cannot call tools can write an action’s name in square brackets around its input, inside its answer. Sent to the system, the tag runs that action.
A link, for an AI in a browser
Some chat AIs can only open links. A token link runs one named action when it is opened, with no key.
A model’s function call, sent as written
# POST https://miscsubjects.com/api/tools/call (header x-sheets-key: YOUR_KEY)
# the body is the tool call exactly as the model wrote it, in any of these shapes:
# OpenAI, Workers AI
{"type": "function", "function": {"name": "NOW", "arguments": "{}"}}
# Anthropic
{"type": "tool_use", "name": "NOW", "input": {}}
# Gemini
{"functionCall": {"name": "NOW", "args": {}}}
# MCP
{"name": "NOW", "arguments": {}}A tag is sent to POST /api/tag as {"text": "[NOW]{}[/NOW]"}. The browser link is /web/run/<action>?args=…&token=….
The record
Every call an AI makes is a Ledger entry
What it asked for, the inputs it chose, what came back, which model, and when.
Live, read from this site just now.
- Check what an AI did before trusting it with more.
- Compare AI models on cost and results, from the record rather than a demo.
- Settle what happened with the record, not with the AI’s own account of it.
Connect your AI app.
Text us which AI app you use and what it should be able to do. You get a key from us, and help connecting it.