Skip to content

modelsalt/model-salt

v1.0.2MIT

An MCP server that gives your model randomness from outside it, so one prompt gives different answers that still do what you asked.

GitHubOpen Website

Changelog

Unreleased

  • random.org is no longer used. Its free terms cover personal, non-commercial use and development only. The physical source is now LfD's quantum random number service; the NIST Randomness Beacon is mixed in next to drand, hashed with local bytes and asked at most once a minute.

1.0.3

  • Keep what you like. Say what to keep ("keep the palette and layout, give me new illustrations") and your model sends those decided answers in the new keep field. Only the open choices are rolled again, the reply starts with a Kept: line that carries over to later runs until you start fresh, and a kept answer that no longer fits the request is dropped and named. Nothing is stored. In a model-fit test Claude Opus 5.5 and GPT-6.1 Sol filled keep when asked to keep something and left it out on "something new", 21 of 21 runs each.

1.0.2

  • Works without Claude Code or Codex. With neither CLI on the machine (a Docker container, Cursor or Copilot alone), the tool no longer returns "nothing was settled" every time: your model does the exploring step itself, in two calls of the tool (the request, then its reply with the id the first call gave). With a CLI, nothing changes. SALT_IN_CONTEXT=1 forces the two calls, 0 the CLI.
  • Every install route on 1.0.2: the Claude plugin pin, server.json (official MCP registry), the MCPB desktop bundle and the root plugin.json. The skill and the bundle's tool text now describe the second call.

1.0.1

  • Listed in the official MCP registry as com.modelsalt/model-salt (mcpName in package.json, server.json). No change to the server.

1.0.0

The first public release.

  • One tool, model_salt, for Claude Code, Codex, Gemini CLI and Kimi Code, with one setup command per client and uninstall to take it out.
  • Tested on six models: Claude Opus 5.5, Claude Fable 5.1, GPT-6.1 Sol, GPT-6 Astra, Gemini 4 Argon and Kimi K3.
  • A randomness setting from 0 to 2 for when you ask for more or less of it. Without one it runs at 1.
  • Two services: anthropic for Claude models (and for now Gemini and Kimi), openai for GPT models.