Setting the model for nicknisi/consultant's gpt5-server MCP

consultant-model-config.md

Setting the model for nicknisi/consultant's gpt5-server MCP

The gpt5-server MCP in the nicknisi/consultant plugin does not read a model from an environment variable. It accepts model as a per-call argument, with 'gpt-5' hardcoded as the Zod default and the || fallback in callGPT5.

So you set it one of three ways.

1. Per call (cleanest)

Pass model in the tool args when invoking it directly:

mcp__plugin_consultant_gpt5-server__gpt5_generate({
  input: "...",
  model: "gpt-5-mini",        // or any GPT-5 variant
  reasoning_effort: "high",
})

This does not go in settings.json — settings.json controls permissions, env vars, hooks, and which MCP servers run, but it can't inject arguments into MCP tool calls. Args are decided by the caller (you, an agent prompt, or a skill).

The durable equivalents:

• In chat: "use the consultant with model: gpt-5-mini" — the assistant passes it through.
• In CLAUDE.md (project or ~/.claude/CLAUDE.md): add a standing instruction such as "When calling mcp__plugin_consultant_gpt5-server__gpt5_generate, always pass model: 'gpt-5-mini' unless I say otherwise." That sticks across sessions.

2. Tell the agent/skill to pass it

The agent prompt fills in the args. Edit ~/.claude/plugins/marketplaces/nicknisi/plugins/consultant/agents/<researcher>.md and add an instruction along the lines of "Always call gpt5_generate with model: 'gpt-5-mini'."

Important caveat about codex-researcher

The codex-researcher agent does not route through gpt5-server. It calls mcp__codex__codex_generate, which is a different MCP server — the codex CLI entry in the plugin's .mcp.json:

"codex": {
  "command": "codex",
  "args": [
    "-m", "gpt-5-codex",
    "-c", "model_reasoning_effort=\"high\"",
    "mcp-server"
  ]
}

For that one, change the -m gpt-5-codex flag in .mcp.json, not anything in gpt5-server.

3. Change the default in the server

Edit either:

• utils.ts:58 — options.model || 'gpt-5'
• index.ts — the Zod .default('gpt-5')

Then rebuild inside mcp-server/ with npm run build.

⚠️ This gets clobbered on plugin update. Prefer options 1 or 2.

TL;DR

Where you want the choice to live	Do this
One-off	Pass model: in the tool call
Across all my sessions	Add a line to ~/.claude/CLAUDE.md
Whenever a specific researcher agent runs	Edit that agent's .md prompt
For the codex server (not gpt5-server)	Change -m in plugin .mcp.json
At the server default level	Edit utils.ts / index.ts, rebuild (fragile)