chatgpt-mcp/docs/SETUP.md

Local Setup Helper — Documentation

Overview

The setup helper is a small interactive onboarding tool for the ChatGPT MCP Server. It reduces configuration friction when choosing between the three supported providers: openai, manual, and ollama.

Run it with:

npm run setup

What it does

1. Asks you to choose a provider — OpenAI, Manual Export, or Ollama
2. Prompts for provider-specific settings — API keys, model names, URLs, etc.
3. Writes a clean .env file — with your selected configuration (asks confirmation first)
4. Optionally creates .claude/settings.local.json — so Claude Code discovers the MCP server automatically

What it does NOT do

• No network requests of any kind
• No OpenAI API key validation against remote servers
• No Ollama model auto-discovery or health checking
• No config migration from previous formats
• No changes to global Claude Code settings (only project-local)
• No external dependencies beyond Node.js built-ins
• No secrets printed to console

Provider selection guide

Provider	Best for	Requires API key?	Cost
openai	Automated second-opinion queries with best-in-class reasoning	Yes (OPENAI_API_KEY)	~$0.003–$0.01 per call
manual	Zero-cost, manual copy-paste workflow	No	$0
ollama	Local/private AI via Ollama's /api/chat endpoint	No	$0 (local compute)

See TASK 10.0 validation report for a detailed comparison of output quality, speed, and trade-offs.

How it works

.env generation

The helper generates a clean .env file with only the keys relevant to your chosen provider:

• CHATGPT_MCP_PROVIDER=<selected> — always written
• OPENAI_API_KEY, OPENAI_MODEL — for openai provider only
• OLLAMA_BASE_URL, OLLAMA_MODEL, OLLAMA_TEMPERATURE, OLLAMA_TIMEOUT — for ollama provider only

If an existing .env file is detected, non-conflicting keys are preserved in a separate section. Provider-specific keys from the old file are not carried over.

.claude/settings.local.json generation

Optionally creates (or overwrites) a project-local Claude Code discovery config:

{
  "mcpServers": {
    "chatgpt-mcp": {
      "command": "npm",
      "args": ["start"]
    }
  }
}

You will be prompted before any file is created or overwritten. Existing files that are not settings.local.json (e.g., other MCP servers) are never modified.

Safety rules

1. No secrets printed. API key input uses terminal masking; if displayed, it shows as sk-***.
2. Confirmation required. You must explicitly confirm before any file is written.
3. No network calls. The helper only reads/writes local files and prompts for input.
4. No global settings changes. Only .claude/settings.local.json (project-local).
5. Git-safe. Both .env and .claude/ are in .gitignore — nothing is committed.

Troubleshooting

"Setup helper requires an interactive terminal"

The helper detects whether stdin is a TTY. If you're piping or running in CI, configure manually via .env:

echo "CHATGPT_MCP_PROVIDER=openai" >> .env
echo "OPENAI_API_KEY=sk-your-key" >> .env

I want to change providers later

Either edit .env directly:

nano .env  # change CHATGPT_MCP_PROVIDER and the provider-specific keys

Or re-run npm run setup — it will generate a new clean .env.

My Ollama model isn't found

The default Ollama model is qwen3:latest (alias for qwen3.6:35b-a3b). If yours isn't available, check with:

ollama list

Then set OLLAMA_MODEL in .env to one of your installed models.

The setup script was accidentally interrupted

No changes are written until you confirm at the end. If only partial writes happened, restore from git:

git checkout -- .env .claude/

Manual configuration reference

If you prefer to configure manually without the helper, create or edit .env with:

OpenAI

CHATGPT_MCP_PROVIDER=openai
OPENAI_API_KEY=sk-your-key-here
OPENAI_MODEL=gpt-5.1
OPENAI_TEMPERATURE=0.2
OPENAI_MAX_OUTPUT_TOKENS=2000

Manual Export

CHATGPT_MCP_PROVIDER=manual

Ollama

CHATGPT_MCP_PROVIDER=ollama
OLLAMA_BASE_URL=http://localhost:11434
OLLAMA_MODEL=qwen3:latest
OLLAMA_TEMPERATURE=0.2
OLLAMA_TIMEOUT=60

---

This helper is intentionally small (one file, zero dependencies). See the TASK 10.2 design for the full specification.