Connect Claude Code to the world: MCP servers (NotebookLM, Supabase, Firecrawl, Playwright), alternative and free models (CCR, Ollama), cloud Routines, and usage monitors.
claude mcp add · /mcp · servers
Supabase · Firecrawl · Playwright
CCR · Ollama · routers
Cloud · /loop · scheduled
tokens · agents · terminals
Sonnet · Opus · Gemini embed
O MCP (Model Context Protocol) is the “USB standard” that connects Claude Code to external tools: databases, browsers, APIs, docs. Each MCP server exposes tools Claude can then use — read from Supabase, scrape a site, control the browser.
Without MCP, Claude can only see your files. With MCP, it acts in the world: queries your actual database, opens pages, and gets live data. It’s what turns the agent from a "code editor" into a "systems operator."
claude mcp add firecrawl "npx -y firecrawl-mcp" # adiciona um server claude mcp list # lista os configurados /mcp # painel dos servers na sessão
| Scope | Where it applies |
|---|---|
| local | Just you, just this project |
| project | Go to .mcp.json (commit, the whole team) |
| user | You, across all projects |
A handful of MCP servers covers most use cases: Supabase (database), Firecrawl (scrape the web), Playwright (control the browser) and NotebookLM (knowledge base). They all come with a claude mcp add.
With Supabase MCP, Claude reads and migrates your real database without you copying the schema; with Firecrawl, it fetches content from any website; with Playwright, it tests the UI by actually clicking through it. That's productivity you can't fake with a prompt.
| MCP | For | Add |
|---|---|---|
| Supabase | Database, auth, migrations | @supabase/mcp-server-supabase |
| Firecrawl | Scrape / index websites | firecrawl-mcp |
| Playwright | Browser, E2E testing | @playwright/mcp |
| NotebookLM | Knowledge base | via the community server |
claude mcp add playwright "npx -y @playwright/mcp@latest" claude mcp add supabase "npx -y @supabase/mcp-server-supabase" -e SUPABASE_ACCESS_TOKEN=...
Tip: after adding, run /mcp to confirm the server connected (green ✓) before asking for the task.
Claude Code looks for the key in ANTHROPIC_API_KEY and the URL in ANTHROPIC_BASE_URL. Replace this URL with a router (CCR — claude-code-router) or a local server (Ollama), you can run it with other models — including for free.
For studying, testing, or simple tasks, you don’t need to spend expensive tokens. CCR routes requests to free or low-cost providers; Ollama runs models on your machine, offline. It keeps the Claude Code workflow and only swaps the “engine.”
npm install -g @musistudio/claude-code-router ccr code # abre o Claude Code roteado p/ outro provedor # github.com/musistudio/claude-code-router
ollama pull qwen2.5-coder export ANTHROPIC_BASE_URL="http://localhost:11434" claude # usa o modelo local, offline
💡 When to use: local/free models are great for learning and light tasks; for serious work, official Sonnet/Opus models deliver much more. Use each where it fits.
The Routines are automations that run in the cloud — without keeping your PC on. You define a prompt, a GitHub repository, a model, and a schedule; the trigger can be scheduling, API or GitHub event. The /loop and the scheduled tasks run locally.
It’s Claude becoming a team member who works independently: reviews PRs, updates dependencies, generates reports—on a schedule, without you. The catch: the cloud environment is stateless e can't see your .env local; secrets go in the cloud environment variables.
| Type | Where it runs | Detail |
|---|---|---|
| Routines | Cloud (stateless) | PC off · minimum cadence ~1h · requires a GitHub repo |
| Scheduled tasks | Local (desktop) | Accesses local files |
| /loop | Local (session) | Short intervals · doesn't survive a restart |
Plan limits: Pro ~5/day · Max ~15/day · Team/Enterprise ~25/day. Each cloud run ≈ 4 vCPU / 16 GB RAM.
⚠️ Gotcha: remote routine doesn't have cookies/a session for your local browser. Automation that depends on a saved login fails — use an API or dedicated authentication, and test multiple times before “production.”
Tools that show, in real time, how much you’re spending e what your agents are doing: token usage, cost, open sessions, and multiple terminals. The community has built several — Token-Dashboard, iccmonit, SMA (Surveil My Agents), Agent View.
Anyone running several agents in parallel (track 6) can lose track of costs and status. A monitor prevents billing surprises and shows which terminal is stuck or finished. /cost handles the basics; dashboards give you the big picture.
Token usage and cost per session/day
Claude Code monitor running on the machine
Monitors several agents at once
Managing multiple terminals/threads
Anthropic releases new versions of Sonnet e Opus frequently — Claude Code points to the latest ones. To semantic search (RAG, "second brain"), the embeddings, where models like Gemini Embedding are popular because they’re cheap/free.
Know the difference between chat model (generates code/text) and embedding model (turns text into vectors for search) avoids confusion. You can switch the first one with /model; the second one feeds your RAG (track 6).
| Type | Does | Examples |
|---|---|---|
| Chat / code | Reasons, edits, chats | Opus · Sonnet · Haiku |
| Embedding | Vectorizes text for search | Gemini Embedding 2.0 |
/model # ver e trocar o modelo de chat ativo /model sonnet # último Sonnet (dia-a-dia) /model opus # último Opus (tarefas difíceis)