Connect Cruxes to Claude Code in two minutes

Claude Code can ask you the same question twice in two sessions. Cruxes keeps the decisions you made in the web app, and the hosted MCP server hands them to the agent, so it starts from what is already settled.

Before you start

• A Cruxes account. Cruxes is in early access: sign in with Google or an email link to get 15 free Cruxes.
• An API key. Create one on the API keys page while signed in. A key is shown once, carries only the scopes you give it (read, resolve, detect), and defaults to read.

Register the server

Put the key in your environment, then register the server with Claude Code over HTTP.

export CRUXES_API_KEY=cxk_...
claude mcp add --transport http cruxes "https://www.cruxes.io/api/mcp" \
  --header "Authorization: Bearer $CRUXES_API_KEY"

Make a first call that costs nothing

Listing tools is free, so it is the right first call. It also proves the key and the header are right before you hand the server to an agent.

curl -s https://www.cruxes.io/api/mcp \
  -H "Authorization: Bearer $CRUXES_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Start from what you decided

In Claude Code, ask it to call get_handoff. It returns the cruxes you decided, each with the option you chose, any that are still open, and the plan. Decided cruxes are settled: the agent should not ask them again, and resolve_crux is the only tool that records a choice.

Reading is free. detect_cruxes and gate_cruxes are metered against your plan the same way the web app is, so an agent that calls them in a loop spends your allowance.

Where next

• The developer guide lists every tool, endpoint and error, with samples in curl, TypeScript and Python.
• The release log shows what changed and when, in plain words.