Model Advisor MCP
# š§ Model Advisor
**Stop wasting your Claude usage. Get a one-line tip on which model to use for each task.**
Using the most powerful model for everything burns through your limits fast. A translation doesn't need Opus, and a full product spec shouldn't be done on Haiku. Model Advisor tells you which one fits, at the start of a task, and tells you at the end if a cheaper model would have done the job.
Free and open source. Built by [SigmaMax AI Solutions](https://sigmamaxaisolutions.com).
> š” **Honest note:** Claude can't switch models for you. Model Advisor gives the advice; you pick the model in Claude's model selector.
| Model | Best for |
|---|---|
| **Haiku 5.5** | Quick and short: rephrase, translate, format, quick facts |
| **Sonnet 5.5** *(default)* | Everyday work: emails, outreach, proposals, research, calculations, normal coding |
| **Opus 5.5** | Complex work: build specs, architecture, strategy, contracts, hard debugging |
| **Fable 5.1** | Only the hardest problems, after Opus falls short |
---
## š Start here: pick your path
| You use... | Use this | Setup time |
|---|---|---|
| **Claude.ai** (web or phone) | **Path 1: add as a connector** (easiest) | 1 minute, no install |
| **Claude.ai** (free plan) | **Path 2: paste instructions** | 2 minutes, nothing to install |
| **Claude Desktop, Claude Code, Antigravity, Cursor** | **Path 3: local MCP server** | 10 minutes |
### Path 1: Add as a connector (easiest, works on phone too)
1. Open **claude.ai** in a browser and go to **Settings ā Connectors ā Add custom connector**.
2. Name: `Model Advisor`. URL: `https://model-advisor-mcp.vercel.app/mcp`
3. Click **Add**, then switch it on from the tools menu in a chat.
4. Try: "Use recommend_model: translate this WhatsApp line into Arabic. I'm on Opus."
Add it once on the website and it appears in the Claude phone app on the same account.
Note: custom connectors may need a paid Claude plan. On a free plan, use Path 2.
### Path 2: Paste instructions (Claude.ai, no install)
1. In Claude, create a **Project** (or open Settings and find your personal preferences).
2. Paste this into the project instructions:
```
At the start of each new task, add ONE line at the top of your reply recommending the best Claude model for it:
"š§ Model: <current> fits ā
" or "š§ Model tip: switch to <model> ā <reason, 10 words max>".
Guide: Haiku = quick/short/mechanical (translate, rephrase, format). Sonnet = default for everyday work (emails, research, calculations, normal coding). Opus = complex multi-step reasoning, build specs, architecture, strategy, contracts, hard debugging. Fable = only the hardest problems after Opus fell short.
At the end of a big task, add "š§ Next time: <model> would be enough" only if a lighter model clearly would have done. If the chat is very long and the topic changed, suggest a new chat. Skip all of this for quick follow-ups.
```
3. Start any chat inside that project. Done.
Prefer a Claude *skill*? The same thing is packaged in [`skills/model-advisor/SKILL.md`](skills/model-advisor/SKILL.md).
### Path 3: Local MCP server (Claude Desktop, Claude Code, Antigravity)
Needs [Node.js 18+](https://nodejs.org) and [Git](https://git-scm.com).
```bash
git clone https://github.com/ahsanmassgazal-cmd/model-advisor-mcp.git
cd model-advisor-mcp
npm install
npm test # should print: 12/12 recommendation cases passed
```
Add it to your MCP config. For **Claude Desktop**, open the config file:
- Windows: press `Win+R`, type `%APPDATA%\Claude`, open `claude_desktop_config.json`
- Mac: `~/Library/Application Support/Claude/claude_desktop_config.json`
```json
{
"mcpServers": {
"model-advisor": {
"command": "node",
"args": ["C:\\Users\\YOU\\model-advisor-mcp\\src\\index.js"]
}
}
}
```
Use the real full path to `src/index.js`. On Windows, double every backslash (`\\`). On Mac, use `/Users/you/model-advisor-mcp/src/index.js`. If the file already lists other servers, add `model-advisor` inside the existing `mcpServers` block.
Fully quit and reopen Claude Desktop. **Claude Code:** `claude mcp add model-advisor -- node /full/path/to/src/index.js`. **Antigravity / Cursor:** add the same `node` command and path in the MCP settings.
**Make it automatic:** add this to your custom instructions so Claude calls it by itself:
> At the start of each new task, call `recommend_model` and show its `one_liner`. At the end of a substantial task, call `review_model_choice` and show the note only if it suggests a change.
---
## š§Ŗ Try it: examples
Just ask in plain words. Real outputs:
| You say | Model Advisor replies |
|---|---|
| *"Use recommend_model: translate this WhatsApp line into Arabic"* (on Opus) | š§ Model tip: Switch down to **Haiku 5.5**, Opus 5.5 is more than this needs (saves usage). |
| *"Draft a LinkedIn outreach message to a logistics CEO"* (on Opus) | š§ Model tip: Switch down to **Sonnet 5.5**, Opus 5.5 is more than this needs (saves usage). |
| *"Write a comprehensive build spec for a multi-agent CRM product"* (on Sonnet) | š§ Model tip: Switch up to **Opus 5.5**, Sonnet 5.5 may give weaker results here. |
| *"Review this founders agreement and flag risky clauses"* (on Haiku) | š§ Model tip: Switch up to **Opus 5.5**, Haiku 5.5 may give weaker results here. |
| *End of a task: "Use review_model_choice: translated one line, used Opus"* | š§ Next time: **Haiku 5.5** would be enough for this kind of task (saves usage). |
| A chat that has run 60+ messages | Start a new chat for a new topic to save tokens. |
## š§ Tools (Path 1 and 3)
| Tool | When | What you get |
|---|---|---|
| `recommend_model` | Start of a task | Best model, a one-line verdict, confidence, tips |
| `review_model_choice` | End of a task | Whether a lighter or stronger model fits next time |
| `list_models` | Any time | The model guide with API ids |
## ā Troubleshooting
- **Connector fails to add:** open the URL in a browser. If you do not see 'Model Advisor MCP is running', the server is down.
- **Tool doesn't appear in Claude Desktop:** the config path or JSON is wrong. Check for single backslashes, missing commas, then fully quit (system tray) and reopen.
- **`node` not recognized:** install Node.js LTS, then reopen your terminal.
- **Web or phone Claude:** use Path 1 (connector) or Path 2 (instructions).
## How it decides
1. Defaults to **Sonnet**, the best balance for most work.
2. Goes **down to Haiku** only for short, mechanical tasks.
3. Goes **up to Opus** for multi-step reasoning, long precise deliverables, or high-stakes work.
4. Suggests **Fable** only when Opus has been tried or the problem is at the frontier.
5. Runs locally with simple rules: **zero API calls, zero tokens**. Edit `MODELS` and `RULES` in [`src/advisor.js`](src/advisor.js) to tune it, and add a case in [`test/test.js`](test/test.js) for each change.
## Roadmap
- Usage log with a weekly "what you could have saved" summary
- Optional LLM-assisted classification for unclear tasks
- Team rule presets
## License
MIT Ā© SigmaMax AI Solutions
---
**Need AI agents for your business?** SigmaMax builds AI agents and automation for companies across the GCC, India and beyond.
š [sigmamaxaisolutions.com](https://sigmamaxaisolutions.com) Ā· āļø ahsansigmamaxai@gmail.com Ā· WhatsApp +968 7694 1459
TDQS
Scored across 3 tools
Each tool occupies a distinct point in the task lifecycle: recommend_model at the start, review_model_choice at the end, and list_models as a static reference lookup. The descriptions explicitly state when to call each, leaving no realistic chance of misselection.
All three names follow the same verb_noun convention (recommend_model, review_model_choice, list_models) with snake_case throughout. The verbs are also semantically aligned with their described actions.
Three tools is on the lean side, but the domain is narrow (model selection advice) and each tool earns its place. Nothing feels padded or redundant.
The surface covers the advisory loop: reference guide, pre-task recommendation, and post-task review. Minor gaps exist, such as comparing two models head-to-head or retrieving past recommendations, but agents can work around these with list_models.