mcp-generator
Generate minimal MCP servers with only the tools your project needs. Reduces context bloat by 80-95%.
Open source Open in the app JSON README (API)
About
Generate minimal MCP servers with only the tools your project needs. Reduces context bloat by 80-95%.
Details
- Kind
- Plugins
- Topic
- No topic detected
- Publisher
- mossae
- Origin
- gemini
- Category
- ferramentas
- Version
- 0.1.0
- Last push
- 2026-03-30T22:40:44Z
- Repository state
- ativo
- Language
- TypeScript
- Added
- 2026-08-30 14:13:39
- Updated
- 2026-08-30 14:13:39
- Origin id
mossae/mcp-generator
README
# MCP Generator
Generate minimal MCP servers for Rocket.Chat with only the tools your project needs. Reduces context bloat by 80-95%.
## The Problem
Existing MCP servers expose all 547 Rocket.Chat API endpoints, burning ~115k tokens per agent loop. Most projects need 2-10 tools. Every unused tool definition wastes tokens and money on every agentic iteration.
## The Solution
This generator produces production-ready MCP servers covering **only** the workflows you select. Generated tools contain real decision logic — not thin API wrappers.
## Quick Start
```bash
npm install
npx tsx src/index.ts generate --capabilities "onboard new team members" --output ./my-server
```
Or with explicit workflow IDs:
```bash
npx tsx src/index.ts generate --workflows "minimal-chatbot,cicd-notifier" --output ./my-server
```
Or with a config file (`.mcp-generator.yaml`):
```yaml
provider: rocketchat
capabilities:
- onboard new members
- CI/CD notifications
output: ./my-server
name: my-rc-mcp
```
```bash
npx tsx src/index.ts generate
```
## CLI Commands
| Command | Description |
|---------|-------------|
| `generate` | Generate a minimal MCP server |
| `list` | Show all available workflows with token costs |
| `measure --all` | Token savings comparison table |
| `measure --capabilities "..."` | Measure savings for natural language input |
## Available Workflows
| Workflow | Steps | Tokens | Savings | Features |
|----------|-------|--------|---------|----------|
| onboard-team-member | 5 | ~184 | 88.8% | decision logic |
| customer-support-bot | 5 | ~163 | 90.1% | decision, events, retry |
| cicd-notifier | 5 | ~162 | 90.2% | decision, retry |
| team-standup | 5 | ~137 | 91.7% | retry |
| content-moderation | 6 | ~138 | 91.6% | decision, events |
| channel-management | 5 | ~145 | 91.2% | decision |
| message-search | 3 | ~128 | 92.2% | |
| file-sharing | 3 | ~120 | 92.7% | |
| analytics-reporter | 4 | ~108 | 93.4% | decision, retry |
| webhook-integration | 4 | ~139 | 91.6% | |
| notification-bot | 2 | ~126 | 92.4% | decision |
| minimal-chatbot | 3 | ~98 | 94.1% | retry |
## What Makes This Different
Other generators produce thin API wrappers:
```typescript
// Thin wrapper
server.tool("onboard_user", {}, async ({ username, channels }) => {
await client.post("/api/v1/users.create", { username });
for (const ch of channels) {
await client.post("/api/v1/channels.invite", { roomId: ch });
}
});
```
This generator produces platform-level tools with real logic:
```typescript
// Generated: decision logic, error handling, rollback
server.tool("rc_onboard_team_member", {}, async ({ username, email, role }) => {
const existing = await client.get("/api/v1/users.info", { username });
if (existing?.user != null) {
userId = existing.user._id;
} else {
const created = await client.post("/api/v1/users.create", {
username, email, password: crypto.randomUUID().slice(0, 16),
});
userId = created.user._id;
}
// Role-based channel selection, per-channel error handling,
// welcome DM with skip-on-failure strategy
});
```
## Architecture
```
User input (NL / config / flags)
│
├─ [NL Parser] ── fuse.js fuzzy matching → workflow templates
│
├─ [ProviderSpec] ── generic platform interface (RC first impl)
│
├─ [WorkflowTemplate] ── decision points, error handlers, rollback
│
├─ [Code Emitter] ── Handlebars → TypeScript with real if/else/try/catch
│
├─ [Token Counter] ── measures and reports savings
│
└─ Generated Output: MCP Server + RC App (if events needed)
```
## Generated Output
Single workflow (no events):
```
my-server/
server/src/
index.ts # MCP server entry
client.ts # RC HTTP client
tools/
minimal-chatbot.ts
package.json
gemini-extension.json
.env.example
```
Workflows with real-time events (monorepo):
```
my-server/
server/ # MCP server
rc-app/ # Rocket.Chat App event bridge
app.json
src/index.ts # Listens for events, forwards via HTTP callback
package.json
gemini-extension.json
```
## gemini-cli Integration
Install generated server as a gemini-cli extension:
```bash
gemini extensions install ./my-server
```
Or configure manually in `~/.gemini/settings.json`:
```json
{
"mcpServers": {
"rocketchat": {
"command": "node",
"args": ["./my-server/server/dist/index.js"]
}
}
}
```
## Development
```bash
npm install
npm test # 88 tests
npm run test:workflows # 54 workflow tests
npm run benchmark # token savings report
```
## Token Savings Benchmark
Combo scenarios across a 10-iteration agent loop:
| Scenario | Tools | Tokens | Full | Savings |
|----------|-------|--------|------|---------|
| Messaging only | 1 | ~98 | ~1648 | 94.1% |
| Support team | 2 | ~301 | ~1648 | 81.7% |
| DevOps | 3 | ~427 | ~1648 | 74.1% |
| Team lead | 3 | ~466 | ~1648 | 71.7% |
| Full admin | 4 | ~575 | ~1648 | 65.1% |
Average single-workflow savings: **91.7%**
## Tech Stack
- TypeScript (strict mode)
- Handlebars (deterministic codegen, zero LLM)
- fuse.js (natural language matching)
- commander + @inquirer/prompts (CLI)
- vitest (88 tests)
- @modelcontextprotocol/sdk (in generated output)
- @rocket.chat/apps-engine (in generated RC App)