Model ID Cheatsheet

Stop your AI coding agent from hallucinating outdated model names. This MCP server gives any AI assistant instant access to accurate, up-to-date API model IDs, pricing, and specs for 107 models across 19 providers.

Built in Go. Single 10MB binary. Zero external calls. Sub-millisecond responses. Auto-updated daily.

- model = "gpt-4-turbo"           # Hallucinated - doesn't exist anymore
+ model = "gpt-5.3-codex"         # Correct - verified against official docs

- model = "claude-3-opus-20240229" # Deprecated
+ model = "claude-opus-4-6"        # Current - latest Anthropic flagship

Quick Start

Pick one option below. You'll be up and running in under a minute.

Option A: Claude Code (one command)

claude mcp add --transport sse --scope user model-id-cheatsheet \
  https://universal-model-registry-production.up.railway.app/sse

Verify it works:

claude mcp list
# Should show: model-id-cheatsheet ... Connected

Then start a new Claude Code session and ask: "What's the latest OpenAI model?" - it will use the tools automatically.

Option B: Cursor

Add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "model-id-cheatsheet": {
      "url": "https://universal-model-registry-production.up.railway.app/sse"
    }
  }
}

Restart Cursor to pick up the change.

Option C: Windsurf

Add to Settings > MCP Servers (or edit ~/.codeium/windsurf/mcp_config.json):

{
  "mcpServers": {
    "model-id-cheatsheet": {
      "serverUrl": "https://universal-model-registry-production.up.railway.app/sse"
    }
  }
}

Option D: Codex CLI

Add to ~/.codex/config.toml:

[mcp_servers.model-id-cheatsheet]
command = "uvx"
args = ["mcp-proxy", "--transport", "sse", "https://universal-model-registry-production.up.railway.app/sse"]

Option E: OpenCode

Add to ~/.config/opencode/opencode.json:

{
  "$schema": "https://opencode.ai/config.json",
  "mcp": {
    "model-id-cheatsheet": {
      "type": "remote",
      "url": "https://universal-model-registry-production.up.railway.app/sse"
    }
  }
}

Option F: Any MCP Client

Connect to the SSE endpoint directly (no API key, no auth):

https://universal-model-registry-production.up.railway.app/sse

Or use the Streamable HTTP transport:

https://universal-model-registry-production.up.railway.app/mcp

Verify Your Setup

Once connected, try asking your AI assistant any of these:

• "What's the correct model ID for Claude Opus 4.6?"
• "Is gpt-4o still available?"
• "Compare gpt-5.2 vs claude-opus-4-6"
• "What's the cheapest model with vision?"

If the agent calls a tool like get_model_info or check_model_status before answering, it's working.

How It Works

Your AI agent gains 6 tools that it calls automatically before writing any model ID:

Resources

What Happens Under the Hood

1. You ask your agent to write code or answer a model question
2. The agent automatically calls the appropriate tool (e.g., get_model_info)
3. The server responds in sub-milliseconds with verified data (no external API calls)
4. The agent writes code with the correct, current model ID

The server instructions tell the agent: "NEVER use a model ID from your training data without verifying it first." This means the agent will always check before writing.

Real-World Examples

Writing an API call:

# You: "Call the OpenAI API with their best coding model"
# Agent calls: get_model_info("gpt-5.4")
response = client.chat.completions.create(
    model="gpt-5.4",  # Verified via model registry
    messages=[...]
)

Catching deprecated models:

# You: "Use gpt-4o for this task"
# Agent calls: check_model_status("gpt-4o")
# Agent: "gpt-4o is deprecated. I'll use gpt-5 instead."
response = client.chat.completions.create(
    model="gpt-5",  # Updated automatically
    messages=[...]
)

Finding the cheapest option:

# You: "Use the cheapest model that supports vision"
# Agent calls: list_models(capability="vision", status="current")
response = client.chat.completions.create(
    model="gpt-5-nano",  # $0.05/$0.40 per 1M tokens
    messages=[...]
)

Comparing options:

# You: "Should I use Claude or GPT for this?"
# Agent calls: compare_models(["claude-opus-4-6", "gpt-5.2"])
# Agent gets a side-by-side table and makes a recommendation

Resource Footprint

A common concern: "Will this slow down my agent or eat tokens?"

For comparison, a single web search costs more tokens than all 6 tool schemas combined.

Covered Models (107 total)

Current Models (79)

Legacy & Deprecated Models (30)

Tracked so your agent can detect outdated model IDs and suggest current replacements:

• OpenAI: gpt-5.3-codex (deprecated), gpt-5.2-codex (deprecated), gpt-5.1-codex-mini (deprecated), o3-pro (deprecated), o3-deep-research (deprecated), o3-mini (legacy), gpt-4.1 (deprecated), gpt-4o (deprecated), gpt-4o-mini (deprecated)
• Anthropic: claude-opus-4-5 (legacy), claude-opus-4-1 (legacy), claude-opus-4-0 (legacy), claude-sonnet-4-0 (legacy), claude-3-7-sonnet-20250219 (deprecated)
• Google: gemini-3-pro-preview (deprecated), gemini-3-pro-image-preview (deprecated), gemini-2.5-flash-lite (deprecated), gemini-2.0-flash-lite (deprecated), gemini-2.0-flash (deprecated)
• xAI: grok-4.1 (deprecated), grok-3 (legacy), grok-3-mini (legacy)
• Mistral: mistral-small-2503 (legacy), codestral-2508 (legacy)
• MiniMax: minimax-m2.1 (legacy), minimax-01 (deprecated)
• Meta: llama-3.3-70b (legacy)
• DeepSeek: deepseek-r1 (legacy), deepseek-v3 (deprecated)
• Zhipu: glm-4.6v (deprecated)

Self-Hosting

If you prefer to run the server locally instead of using the hosted endpoint:

Option 1: Build from Source (recommended for local use)

Requires Go 1.23+.

git clone https://github.com/aezizhu/universal-model-registry.git
cd universal-model-registry/go-server
go build -o model-id-cheatsheet ./cmd/server

Then add it to Claude Code as a local stdio server (zero latency, no network):

claude mcp add --scope user model-id-cheatsheet -- /path/to/model-id-cheatsheet

Or run in SSE mode for other clients:

MCP_TRANSPORT=sse PORT=8000 ./model-id-cheatsheet
# Endpoint: http://localhost:8000/sse

Option 2: Docker

git clone https://github.com/aezizhu/universal-model-registry.git
cd universal-model-registry
docker build -t model-id-cheatsheet .
docker run -p 8000:8000 model-id-cheatsheet

Your SSE endpoint will be at http://localhost:8000/sse.

Option 3: Deploy to Railway

Or manually:

railway login
railway init
railway up

Staying Up to Date

Model data is automatically checked and updated daily at 7 PM Pacific Time -- no human intervention needed.

How it works:

1. Railway cron runs the updater daily, scraping 6 providers' public documentation pages (no API keys needed)
2. Models removed from docs --> auto-deprecated via PR (status changed to "deprecated" in code)
3. New models detected --> GitHub issue created for review
4. CI runs on the auto-generated PR --> if tests pass --> auto-merged into main
5. Railway auto-deploys from main

No provider API keys required. The updater reads publicly available documentation pages to detect model changes. Only GITHUB_TOKEN and GITHUB_REPO are needed for creating PRs and issues.

Railway Cron (primary) -- The hosted instance uses a Railway cron service that runs the updater daily. See configs/railway-updater.toml for the configuration.

Required env vars (set in Railway dashboard):

• GITHUB_TOKEN -- GitHub personal access token with repo scope
• GITHUB_REPO -- Repository in "owner/repo" format (e.g. "aezizhu/universal-model-registry")

Providers checked (via public docs):

• OpenAI (via GitHub SDK source), Anthropic, Google, Mistral, xAI, DeepSeek

CI/CD Workflows:

• .github/workflows/ci.yml -- runs tests on every PR
• .github/workflows/auto-merge.yml -- auto-merges bot PRs (labeled auto-update) after CI passes

GitHub Actions (alternative) -- A GitHub Actions workflow is also included at .github/workflows/auto-update.yml for users who self-host without Railway. No API keys needed -- only GITHUB_TOKEN (automatically provided by GitHub Actions).

Security

• Rate limiting: 60 requests/minute per IP
• Connection limits: Max 5 SSE connections per IP, 100 total
• Request body limit: 64KB max
• Input sanitization: All string inputs truncated to safe lengths
• HTTP hardening: ReadTimeout 15s, ReadHeaderTimeout 5s, IdleTimeout 120s, 64KB max headers
• Non-root Docker: Containers run as unprivileged user
• Graceful shutdown: Clean connection draining on SIGINT/SIGTERM

Tech Stack

• Language: Go 1.23
• MCP SDK: github.com/modelcontextprotocol/go-sdk v1.3.0 (official)
• Transports: stdio, SSE, Streamable HTTP
• Binary size: ~10MB
• Tests: 156 unit tests
• Security: Per-IP rate limiting, connection limits, input sanitization
• Deploy: Docker (alpine), Railway

Contributing

Contributions are welcome! Whether it's adding a new model, fixing data, or improving the server:

1. Fork the repo and clone it locally
2. Edit model data in go-server/internal/models/data.go
3. Update test counts in go-server/internal/models/data_test.go
4. Run the tests:
   bash

   cd go-server && go test ./... -v
   

5. Submit a PR -- we'll review it quickly

If you spot an outdated model or incorrect pricing, opening an issue is just as helpful.

License

MIT