Cerell-mcp
# Cerell MCP — Infrastructure Copilot & Developer Cashback
[](https://modelcontextprotocol.io)
[](https://nodejs.org)
[](LICENSE)
[](https://cerell.eu)
[](https://smithery.ai/servers/apptent/cerell)
## Mission
The decision about which database, hosting, or auth service goes into a project no longer happens in a browser tab -- it happens inside an AI coding agent. Cerell is built for exactly that moment, and it solves two problems at once.
**For developers:** AI-assisted coding is expensive. Running agents through Cursor, Claude, or Windsurf burns real money in LLM tokens -- often $60-100/month for daily users. Cerell subsidizes that cost. When your agent suggests a partner service (a database, an email provider, a cache layer) and you accept it, the partner co-funds the tokens used in your session. You also earn cashback on your monthly cloud spend with that partner, credited back to your Cerell account. You build faster, you spend less.
**For SaaS partners:** Developer tools spend millions acquiring customers who sign up and never activate. Credits are handed out blind. Ads are clicked and forgotten. Cerell flips the model -- partners pay only when their product is actually wired into a developer's repository, after the developer explicitly accepted the suggestion. The conversion event is the **install in the repo**, not a pageview, not a form submission. That is a fundamentally different -- and cheaper -- unit of acquisition.
**How it works in practice:**
1. Your AI agent asks to add a database (or auth, hosting, email, cache...).
2. Cerell surfaces a curated partner match with live pricing and cashback rate.
3. You review, see the alternatives, and accept or skip -- your choice, always.
4. On accept, the upstream partner MCP provisions the resource and returns credentials directly to your environment.
5. The partner is billed for the install. You earn cashback. The session cost is co-funded.
One action. Two problems solved.
## Agenda
- **Close the gap between AI suggestion and real adoption.** MCP registries offer discovery; Cerell adds distribution, attribution, and economics.
- **Make infrastructure decisions transparent.** Every partner suggestion is labeled. Alternatives are always shown. Nothing is installed without the developer accepting.
- **Align incentives across the stack.** Partners pay for activated integrations. Developers earn back a share of what they spend on cloud. The platform takes a margin only when value is delivered.
- **Stay editor-agnostic.** Cursor, Claude Code, Windsurf, VS Code -- the same `infra` tool works everywhere, keeping the acquisition channel neutral and trustworthy for partners who do not want to depend on a single IDE vendor.
- **Grow the catalog without growing complexity.** Partners are added as JSON files in `data/partners/`. The MCP tool surface stays constant -- one tool, every category.
---
Local **stdio** meta-tool gateway for Cursor, Claude Code, Windsurf, and VS Code.
One tool — `infra(mode, request)` — covers compare → provision → connect → operate/optimize/evolve → cashback, while keeping affiliate signup links in the path.
---
## Quickstart
```bash
git clone https://github.com/Pawelec12/cerell-mcp.git
cd cerell-mcp
npm install
npm run build
```
### Cursor (`.cursor/mcp.json`)
```json
{
"mcpServers": {
"cerell": {
"command": "node",
"args": ["C:/path/to/cerell-mcp/dist/index.js"]
}
}
}
```
### Claude Code
```bash
claude mcp add cerell -- node /ABSOLUTE/PATH/TO/cerell-mcp/dist/index.js
```
---
## Tool: `infra`
| Arg | Meaning |
|-----|---------|
| `mode` | `compare` \| `provision` \| `connect` \| `disconnect` \| `status` \| `optimize` \| `evolve` \| `cashback` \| `portal` |
| `request` | Free text / tokens. Flags: `detail=1`, `limit=5` (use `limit=0` for all), `offset=N` |
Replies are **compact JSON** by default (`{"ok":true,...}`). Add `detail=1` for full pricing, actions, docs, migration steps.
### Examples
```text
mode=compare request="database postgres"
mode=compare request="auth detail=1 limit=10"
mode=provision request="neon"
mode=connect request="neon api_key=YOUR_KEY plan=Launch"
mode=provision request="neon create_database name=app region=aws-eu-central-1"
mode=status request=""
mode=optimize request=""
mode=evolve request="from=neon to=supabase detail=1"
mode=cashback request=""
mode=cashback request="amount=25 method=revolut revtag=@you"
mode=portal request=""
```
Flow after you pick a partner: **signup (affiliate URL) → connect API key → provision action → cashback tracking**.
Partner upstream MCPs (optional `upstream` in partner JSON) start **lazily** only when linked and an action needs them (max 3 live connections, LRU).
---
## Partners
Configured in `data/partners/*.json` (Neon, Clerk, Upstash, Resend, Vercel, …). Catalog size does not grow the MCP tool list.
---
## Publish on Smithery (local / MCPB)
Do **not** paste the GitHub URL into Smithery’s “MCP Server URL” field. Cerell is stdio-local.
```bash
npm run pack:mcpb
# → cerell.mcpb
npx @smithery/cli mcp publish ./cerell.mcpb -n apptent/cerell
```
That matches Smithery’s **Local (MCPB Bundle)** flow: users download the bundle and run it on their machine.
## Dev
```bash
npm run build
npm test
npm run pack:mcpb
npm run inspector
```
---
## License
MIT © [Pawelec12](https://github.com/Pawelec12) · [cerell.eu](https://cerell.eu)
TDQS
Scored across 1 tool
With only one tool, there is no cross-tool overlap or misselection risk. However, the single tool bundles many distinct operations (partners, stack, cashback, portal), so agents must disambiguate through parameters rather than tool choice.
There is only one tool name, so consistency cannot be meaningfully evaluated. The name 'infra' is broad and does not follow a verb_noun pattern, but it is readable.
The apparent scope spans partner provisioning, stack optimization, evolution, cashback, and portal operations. Collapsing all of that into one mega-tool is too thin for the domain and forces excessive reliance on parameter conventions.
The description mentions several major capability areas, suggesting broad coverage, but the surface lacks granular operations and clear CRUD/lifecycle boundaries. Agents may work around this with the detail/limit parameters, but specific workflows are opaque.