Connecting an MCP client

Radiant answers MCP, so an assistant running outside Radiant — Claude Code, Codex, anything that speaks the protocol — can read your maps and change them. It acts as you, with the permissions you already have, in every organization you belong to.

This is a different thing from the assistant in the chat panel. That one lives inside a project and follows its permission mode. An MCP client is your own tool, and its changes use the access you grant when you connect it.

Why connect oneCopy Anchor Link

• The context is already in your tool. A client that can see your repository, your notes or a spreadsheet can build a map out of them without you moving anything across by hand.
• Bulk and repeat work. Twenty nodes, a tournament's worth of questions, or the same structure rebuilt across several projects — one call rather than an afternoon of clicking.
• Your choice of model and client. The chat panel offers the models Radiant supports. Over MCP you bring your own, and keep whatever setup you already work in.
• Radiant as one source among several. A client that also reads your other systems can carry a number into a map, or take a simulated result back out, without either side hosting the other.
• Reading a map elsewhere. Asking what a map contains, what its model compiles to, or why a node is blocked works without opening the project.

Most MCP tools also keep working when your AI credits are spent, since a client brings its own model.

ConnectingCopy Anchor Link

Connect your agent in the top bar opens the setup, from anywhere in Radiant.

The short way is Onboard your local agent to Radiant. It copies a one-line prompt; paste it into your agent and it fetches the setup instructions from your Radiant address at /agent-setup/prompt.md, then registers Radiant itself. The instructions are public, so you can read them before you paste. The same button sits under the starter prompts of a new chat.

Set up by hand holds the same pieces for doing it yourself:

• The endpoint — your Radiant address followed by /mcp, a streamable HTTP URL.
• A command for your client, with Claude Code and Codex offered as tabs.
• For an assistant in the browser, a Claude tab with a button that opens Claude's connector dialog filled in for Radiant, and a ChatGPT tab with the steps to add Radiant there yourself.

There is no token to copy either way. The first time the client calls Radiant, it opens your browser: you sign in if you are not already, read what the client asks for, and allow it. The client then holds its own credentials, renews them on its own, and never shows them to you.

What you allowCopy Anchor Link

The connect screen lists the organizations the client will reach, and what it asked for:

• Read your projects and maps
• Create and change projects, nodes, and connections
• Operationalize questions and get AI forecasts using your AI credits
• Read Metaculus questions and forecasts
• Stay connected without signing in again

Allowing does not widen what you can reach. Every call runs as you, so a client sees the projects you can see and can change the ones you can already change. A project you have read-only access to stays read-only over MCP, and a viewer's connection cannot write anywhere.

What a client can doCopy Anchor Link

Reading, without changing anything:

• List your projects, read a map, and query parts of one
• Inspect the forecasting model a map compiles to, including its diagnostics
• Run a simulation and read the result
• Search Metaculus posts and read individual ones
• Search this documentation and read a page from it
• List operationalization and AI forecast tasks and check their status and results

Changing:

• Create a project, and update its title or description
• Add, edit and delete nodes
• Draw, edit and delete connections
• Apply a batch of map changes in one call
• Import Metaculus questions onto a map
• Operationalize a question or get an AI forecast for a node
• Accept a finished operationalization or AI forecast onto its node, or dismiss it

What it cannot doCopy Anchor Link

• No Radiant approval step. Permission modes and approvals belong to the chat panel. Your MCP client may ask for confirmation before it acts, but Radiant does not.
• No chat history, and no review of other tasks. MCP cannot read chats. Of the tasks that wait for your review, it can accept or dismiss only operationalizations and AI forecasts.
• No saved runs. A client can run a simulation to answer a question, but cannot open a run you saved earlier.

DisconnectingCopy Anchor Link

Account settings lists every client you have allowed under Connected apps. Disconnect ends its access in every organization: the client cannot renew its credentials and has to ask again the next time it needs Radiant. A credential it already holds lasts a few more minutes at most. Anything the client already changed stays.

An API key issued before connections existed still works on the endpoint. Keys are listed under API keys in Account settings, where you can revoke them; new connections do not need one.

Telling MCP changes apartCopy Anchor Link

Nodes and connections record how they arrived. One created by a client reads over MCP on the authorship line in the inspector's Info section, next to your name — the same place an agent's edit reads with an agent.

Undo does not cover them, for the same reason it does not cover an agent's edits: the undo stack follows what you did in your browser. Use Versions to go back.

CreditsCopy Anchor Link

Ordinary MCP calls do not draw on your monthly AI credits. Starting question operationalization or an AI forecast does. The tool returns a task ID so your client can check the result later.

An exhausted allowance does not stop ordinary MCP work, but prevents new operationalizations and AI forecasts. When credits run out, AI credits in the top bar offers Connect your agent or Request more credits, which opens the contact page.