MCP server

curva mcp lets any Model Context Protocol client (desktop assistants, coding agents, IDE agents) call Curva as a tool. An agent then gets typed answers with calibrated probabilities instead of judging in free text.

It speaks MCP over stdio (JSON-RPC 2.0, one message per line) and forwards to a running Curva server, so decisions go through that server’s cache, calibration, auth and audit log.

curva serve &                                                   # or use a shared server
curva mcp                                                       # default --url http://127.0.0.1:7777
curva mcp --url https://curva.internal --api-key curva_...       # or set CURVA_API_KEY

Add it to a client

Most MCP clients take a command to launch. For example, with Claude Code:

claude mcp add curva -- curva mcp --url http://127.0.0.1:7777

Or in a JSON client config such as Claude Desktop’s claude_desktop_config.json:

{"mcpServers": {"curva": {"command": "curva", "args": ["mcp"], "env": {"CURVA_API_KEY": "curva_..."}}}}

The decide tool

Argument	Required	Meaning
state	yes	Any JSON to decide about
questions	yes	Key → question, as in /v1/decide (choice, score, noul, multi)
model	no	A model id, a fallback list, {"council": [...]} or {"cascade": [...]}
project	no	Calibration namespace

Other arguments are dropped. The result is the /v1/decide response JSON as text. Any failure (server down, 401, 422, 429, 502) comes back as a tool error (isError: true) whose text holds Curva’s error message, so the agent can fix the request or retry.

Supported methods: initialize, notifications/initialized, ping, tools/list and tools/call. Requests are answered one at a time.