Ten runs, one answer.
Ask a model the same thing again and again and you get the same answer back.




















The fix: add a dash of entropy to your query.
Model Salt is one MCP tool, model_salt. Before an open task the model calls it once, gets its open choices settled by randomness from outside the model, and does the work.
Want more or less? Ask for it: the tool takes a randomness setting from 0 to 2, and 1 is the default. See the settings
npx -y @modelsalt/model-salt setup claude- Run the command above in a terminal. It adds Model Salt to Claude Code for every project.
- Start a new Claude Code session.
- Check:
claude mcp listshows model-salt as connected, or type/mcpin a session. - Ask for something open, like "Suggest a name for a café": the model calls
model_saltbefore it answers.npx -y @modelsalt/model-salt uninstall claudetakes it out.
npx -y @modelsalt/model-salt setup codex- Run the command above in a terminal. It adds Model Salt, lists its tool up front and adds a skill in
~/.agents/skills/that tells the model when to call it. - Start a new Codex session.
- Check:
codex mcp listshows model-salt, or type/mcpin a session. - Ask for something open, like "Suggest a name for a café": the model calls
model_saltbefore it answers.npx -y @modelsalt/model-salt uninstall codextakes it out.
npx -y @modelsalt/model-salt setup gemini- Run the command above in a terminal. It adds Model Salt to Gemini CLI for every project.
- Start Gemini CLI in a folder you trust: it runs MCP servers only there.
- Check: type
/mcp. model-salt is listed as ready, with its one tool. - Ask for something open, like "Suggest a name for a café": the model calls
model_saltbefore it answers.npx -y @modelsalt/model-salt uninstall geminitakes it out.
npx -y @modelsalt/model-salt setup kimi- Run the command above in a terminal. It adds Model Salt and a skill in
~/.agents/skills/that tells the model when to call it. - Start a new Kimi Code session.
- Check: ask for something open, like "Suggest a name for a café": the model calls
model_saltbefore it answers. npx -y @modelsalt/model-salt uninstall kimitakes it out.
{
"mcpServers": {
"model-salt": {
"command": "npx",
"args": ["-y", "@modelsalt/model-salt"]
}
}
}- In Claude Desktop open Settings, Developer, Edit Config. It shows
claude_desktop_config.json(Mac:~/Library/Application Support/Claude/, Windows:%APPDATA%\Claude\). - Paste the block above into the file. If it already has
"mcpServers", add only the"model-salt"entry inside it. Save. - Quit Claude Desktop fully (not just the window) and open it again. It needs Node 18 or later for
npx. - Check: Settings, Developer shows model-salt as running, and the tools menu in a new chat lists it.
- Ask for something open, like "Suggest a name for a café": the model calls
model_saltbefore it answers.
{
"mcpServers": {
"model-salt": {
"type": "stdio",
"command": "npx",
"args": ["-y", "@modelsalt/model-salt"]
}
}
}- Press Add to Cursor above, or paste the block into
~/.cursor/mcp.json(every project) or.cursor/mcp.json(one project). - Check: Cursor Settings, MCP shows model-salt with a green dot and 1 tool.
- In Agent chat: ask for something open, like "Suggest a name for a café": the model calls
model_saltbefore it answers. - Using GPT models? Add
"openai"as the last argument.
code --add-mcp '{"name":"model-salt","type":"stdio","command":"npx","args":["-y","@modelsalt/model-salt"]}'- Press Add to VS Code above, or run the command in a terminal. It adds Model Salt for GitHub Copilot.
- For one project only, put
{"servers":{"model-salt":{"type":"stdio","command":"npx","args":["-y","@modelsalt/model-salt"]}}}in.vscode/mcp.jsoninstead. - Check: Command Palette, MCP: List Servers shows model-salt running; in Copilot Chat, Agent mode, the tools button lists
model_salt. - In Agent mode: ask for something open, like "Suggest a name for a café": the model calls
model_saltbefore it answers. - Using GPT models? Add
"openai"as the last argument.
{
"mcpServers": {
"model-salt": {
"command": "npx",
"args": ["-y", "@modelsalt/model-salt"]
}
}
}- In Windsurf open Cascade, then the MCP servers icon and View raw config (
~/.codeium/windsurf/mcp_config.json). - Paste the block above. If the file already has
"mcpServers", add only the"model-salt"entry. Save, then press Refresh. - Check: model-salt shows in the MCP servers list with its tool on.
- In Cascade: ask for something open, like "Suggest a name for a café": the model calls
model_saltbefore it answers. - Using GPT models? Add
"openai"as the last argument.
cline mcp add model-salt --yes -- npx -y @modelsalt/model-salt- With the Cline CLI, run the command above. In the VS Code extension instead: MCP Servers, Configure, Configure MCP Servers, and add
{"mcpServers":{"model-salt":{"command":"npx","args":["-y","@modelsalt/model-salt"]}}}. - Check: MCP Servers, Installed shows model-salt with a green dot.
- Ask for something open, like "Suggest a name for a café": the model calls
model_saltbefore it answers. - Using GPT models? Add
"openai"as the last argument.
{
"context_servers": {
"model-salt": {
"command": "npx",
"args": ["-y", "@modelsalt/model-salt"],
"env": {}
}
}
}- Open Zed's
settings.json(command palette: zed: open settings file) and add the block above, or use Settings, AI, MCP Servers, Add Local Server. - Check: the Agent Panel's settings show model-salt with a green dot.
- In the Agent Panel: ask for something open, like "Suggest a name for a café": the model calls
model_saltbefore it answers. - Using GPT models? Add
"openai"as the last argument.
{
"mcpServers": {
"model-salt": {
"command": "npx",
"args": ["-y", "@modelsalt/model-salt"]
}
}
}- Any MCP client that runs local (stdio) servers: the command is
npx -y @modelsalt/model-salt. It needs Node 18 or later and no keys. - Add the block above where your client keeps its MCP servers, then restart the client.
- Check: the client's tool list shows
model_salt. - Ask for something open, like "Suggest a name for a café": the model calls
model_saltbefore it answers. - Using GPT models? Add
"openai"as the last argument.
Make an image or other asset in ChatGPT or Gemini with nothing to install. You make one message here and paste it into a chat. The chat settles the open choices with fresh random numbers, shows you its picks, and asks before it generates anything.
- What you get back first: five choices with the odds of each answer, the pick for each from its random number, a Chosen list, and the question "Shall I generate it from these picks?". Nothing is generated yet.
- Reply yes and it generates the asset from those picks. Want a pick changed? Say which, and it asks again.
- Want another take? Press Make the message again for new numbers. Reusing a message gives the same picks.
The same prompts, fixed with Model Salt @1























Temperature vs Model Salt
A test: one model, one prompt, ten runs a row. In the first row only the temperature changes. In the second only the Model Salt dial changes.
- Model
- Gemini 4 Argon
- Prompt
- “Generate an SVG of a pelican riding a bicycle.”
- Values
- 0.2 to 2.0, one run each
0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0
0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0
Temperature makes less likely words more likely, one word at a time. The model still writes its usual answer, so only the sky moves.
Model Salt has the model write its usual answer and the choices that make it, then moves drawn from randomness outside the model change those choices before it does the work. The dial says how far they go: 0 keeps the usual answer, 2 makes every answer equally likely.
Choose how random
Setting 0: Always the usual
Setting 0.5: Mostly the usual
Setting 1: Default
Setting 1.5: Spread wider
Setting 2: All equally likely
Setting 0: Always the usual
Setting 0.5: Mostly the usual
Setting 1: Default
Setting 1.5: Spread wider
Setting 2: All equally likely
Setting 0: Always the usual
Setting 0.5: Mostly the usual
Setting 1: Default
Setting 1.5: Spread wider
Setting 2: All equally likely
Setting 0: Always the usual
Setting 0.5: Mostly the usual
Setting 1: Default
Setting 1.5: Spread wider
Setting 2: All equally likely
Setting 0: Always the usual
Setting 0.5: Mostly the usual
Setting 1: Default
Setting 1.5: Spread wider
Setting 2: All equally likely
Setting 0: Always the usual
Setting 0.5: Mostly the usual
Setting 1: Default
Setting 1.5: Spread wider
Setting 2: All equally likely























































































































