A job-by-job shortlist of the MCP servers worth adding to Claude Code in 2027, from Context7 and GitHub to Playwright, Sentry, and Firecrawl. Each pick lists the install command, what is free, what costs money, and how much context it uses.
Claude Code can already read files, run shell commands, and edit code. What it cannot do alone is read your production errors, open a real browser, pull this week's library docs, or query your database. MCP servers add those abilities, and there are thousands of them, which is exactly the problem. Most roundups list 50 servers. You need about five, and every extra one costs context on every turn.
This shortlist is organized by job, with the install command, the real price, and the context cost of each pick. It was checked in October 2026, so open each pricing page before you commit, because free tiers in this corner of the market change often.
| Server | Best for | Cost | Runs as |
|---|---|---|---|
| Context7 | Current library docs | Free tier, then paid | Local command |
| GitHub | Repos, issues, pull requests | Free with an account | Remote HTTP |
| Playwright | Driving a real browser | Free, open source | Local command |
| Chrome DevTools | Console, network, performance | Free, open source | Local command |
| DBHub (Postgres) | Querying a database | Free, open source | Local command |
| Sentry | Production errors | Account required | Remote HTTP |
| Notion | Specs and notes | Workspace plan | Remote HTTP |
| Firecrawl | Scraping and crawling | Free credits, then paid | Local command |
| PicassoIA | Images and video | Plan dependent | Remote connector |
Most MCP servers are free software. The money goes to whatever sits behind them: a hosted docs index, a scraping service, a ticket tracker, a database. Sort every server into one of three buckets before you install it.
Open source servers such as Playwright MCP, Chrome DevTools MCP, and DBHub cost nothing to run. They use your machine and your existing accounts. Hosted servers such as Context7 and Firecrawl are free up to a quota, then they charge. Connectors for products you already pay for, like Notion, Sentry, Stripe, or Asana, add no fee of their own, but they inherit the limits of the plan behind them.
| Bucket | Examples | What you pay |
|---|---|---|
| Free and local | Playwright, Chrome DevTools, DBHub | Nothing, only your CPU and tokens |
| Free tier, then paid | Context7, Firecrawl | A quota, then a monthly plan |
| Free connector, paid account | Sentry, Notion, Stripe, Asana, HubSpot | Whatever the underlying plan costs |
💡 Check the credential before you paste it. Prefer read-only tokens, test-mode accounts, and one scoped token per project. An MCP server acts with whatever access you hand it, and a prompt injection hidden in a scraped page can try to use that access.
Every connected server advertises its tools to the model, and those descriptions take up room. One third-party comparison measured Playwright MCP at roughly 13.7k tokens and Chrome DevTools MCP at roughly 19k tokens when fully loaded, which is a visible slice of a context window. Claude Code now loads tools lazily by default, so idle servers cost far less than they used to. A short list still wins: fewer servers means fewer wrong tool picks and fewer surprises.
Output is capped as well. Claude Code warns at 10,000 tokens of tool output and allows 25,000 by default. If a database server returns long result sets, raise the cap with the MAX_MCP_OUTPUT_TOKENS environment variable.
Remote servers use HTTP. Local servers run as a command, and the command goes after --:
# Remote server over HTTP
claude mcp add --transport http sentry https://mcp.sentry.dev/mcp
# Local server over stdio
claude mcp add playwright -- npx @playwright/mcp@latest
# Check health, then authenticate inside a session
claude mcp list
/mcp
Where a server is stored decides who gets it:
| Scope | Stored in | Shared with the team | Applies to |
|---|---|---|---|
| Local (default) | ~/.claude.json | No | The current project only |
| Project | .mcp.json in the repo | Yes, through git | The current project only |
| User | ~/.claude.json | No | All your projects |
When the same name appears twice, local beats project and project beats user. Project servers need your approval before they run, and .mcp.json should never hold a secret. Use ${VAR} expansion so the value comes from your environment instead.
If you install only four, make it these. Between them they handle documentation, code hosting, and the browser, which is where coding agents go wrong most often.
Context7, built by Upstash, injects version-specific library documentation into the prompt on demand. It targets the most common failure in AI coding: a confident call to an API that changed two releases ago. It earns its place with fast-moving frameworks such as Next.js or Tailwind. Ask for it by name: "use context7 to check the current routing API."
Cost: a free account with a monthly request quota, then paid plans. Reports from early 2026 put the free quota around 1,000 requests a month after a cut and a partial restore, so confirm the current number. Install: claude mcp add context7 -- npx -y @upstash/context7-mcp, then attach the credential from your Context7 dashboard if the free tier asks for one.
The official GitHub server exposes repositories, issues, pull requests, and code search. The remote endpoint accepts a personal access token:
claude mcp add --transport http github https://api.githubcopilot.com/mcp/ \
--header "Authorization: Bearer YOUR_GITHUB_PAT"
Cost: free with a GitHub account. Create a fine-grained token limited to the repositories you actually work in, with read access unless you want Claude opening pull requests for you. If you already lean on the gh command line tool, the gain is smaller, because Claude Code can run it through bash. The server pays off when you want structured results instead of parsed text.
Microsoft's Playwright MCP drives a real browser through accessibility snapshots rather than screenshots, so the output is text, fast, and cheap to read. It clicks, fills forms, takes screenshots, and works across Chromium, Firefox, and WebKit. A good first prompt: "Open the signup page, submit an empty form, and tell me exactly what the error messages say." It can also turn a flow you just ran into a Playwright test that lives in your repository.
Cost: free and open source. Install: claude mcp add playwright -- npx @playwright/mcp@latest. Expect it to be one of the heavier servers in context.
Google's Chrome DevTools MCP is built for inspecting a page, not driving it. It reads console messages with source-mapped stack traces, lists network requests, and records performance traces with Core Web Vitals. The common pairing is simple: Playwright drives the flow, DevTools measures it.
Cost: free and open source. Install: claude mcp add chrome-devtools -- npx chrome-devtools-mcp@latest.
💡 Context tight? Run one browser server per session. Use Playwright when you are testing behavior and Chrome DevTools when you are chasing a slow page. Running both at once rarely pays for the tokens.
Anthropic's own docs use DBHub from Bytebase as the database example, and it is a sensible default. It gives Claude schema browsing and SQL execution through one local command:
claude mcp add --transport stdio db -- npx -y @bytebase/dbhub \
--dsn "postgresql://readonly:pass@prod.db.com:5432/analytics"
Cost: free and open source. The important part is the credential in that connection string: create a read-only database user, and point it at a replica or a staging copy. A model that can write to your primary database will eventually do so at the worst possible moment.
Sentry's hosted server at mcp.sentry.dev/mcp uses OAuth. Add it, run /mcp, and log in through the browser. Claude can then pull issues, stack traces, and the release that introduced a regression, so "fix the top crash from last night" starts from real data instead of a guess.
Cost: you need a Sentry account. The connector adds no fee, but what you can query depends on your plan.
Notion's hosted server lets Claude read a spec and write the changelog back to the same page:
claude mcp add --transport http notion https://mcp.notion.com/mcp
Authentication runs through OAuth, like Sentry. Cost: the connector is free, and your workspace plan sets the limits.
Firecrawl turns web pages into clean markdown, runs searches, and crawls whole sites. Claude Code already has built-in fetch and search tools, so reach for Firecrawl when a page needs JavaScript rendering or you want an entire documentation site in one pass.
Cost: free credits to start, then paid plans. Recent pricing roundups list an entry plan around $16 a month on annual billing, with a search or a page scrape costing one credit each. Check the current numbers before budgeting a large crawl.
Business tools follow the same pattern as Sentry and Notion. Anthropic's docs show remote servers for Stripe at https://mcp.stripe.com and HubSpot at https://mcp.hubspot.com, plus Asana at https://mcp.asana.com/sse, which uses the older SSE transport that Claude Code still supports but marks as deprecated. Connect them in test mode or a sandbox account first. A billing server that can issue refunds deserves a slow rollout.
The reference Filesystem server shows up on almost every list, but Claude Code already ships with tools to read, write, edit, and search files. Adding a second path to the same disk gives you duplicate tools and a larger context bill. Skip it unless you need controlled access to a directory outside the project.
This reference server structures multi-step reasoning as tool calls. It became popular when models had no built-in thinking mode. With current Claude models that reason natively, the gain is small and the tokens are real. Try your hardest task without it first, and keep it only if the results improve.
Coding sessions often need visuals too: a hero image for a landing page, a social card, a product mockup, a short clip for a changelog. An image and video MCP server lets Claude Code generate them inside the same session, instead of sending you to another tab.
Picasso IA offers a developer API at https://api.picassoia.com/v1 with Bearer token authentication and Replicate-style endpoints. Jobs are asynchronous: you create a prediction, poll it, then fetch the result. The same four models are available over the API and through MCP:
| Model | Job |
|---|---|
| PicassoIA Image | Text to image |
| PicassoIA Image Editor Pro | Editing an existing image |
| PicassoIA Video | Video from text or an image |
| Seedance 2.5 Lite | Video with audio |
Limits worth knowing: 5 concurrent predictions per account, shared across tokens and MCP connections, prompts up to 4,000 characters, a 10 MB request body, and a 3-hour timeout. Which plan includes API access and MCP connections is listed on the pricing page, and I would rather send you there than quote a number that may have moved.
MCP connections are created from picassoia.com/en/mcp/accounts after you log in. Claude Code also lists connectors from your claude.ai account next to the servers you add yourself, with the lowest precedence of any source, so a local server with the same name always wins.
Because jobs are asynchronous and capped at 5 at a time, ask for images in small batches. A prompt that works well: "Generate a 16:9 hero image for the pricing page in a photorealistic style, then add the returned URL to the page frontmatter."
One more place a model helps is a sanity check on your MCP setup before you commit it. Claude Sonnet 5 on Picasso IA is built for coding and tool-use tasks, reads images, and has an adjustable effort level, which suits a config review.
.mcp.json with every token replaced by a placeholder such as YOUR_TOKEN. Never paste a live credential.low, which turns thinking off. Use medium for a quick pass and high when several servers overlap./mcp panel or an error message in the image field, so the model sees what you see.💡 Ask for a ranked "drop first" list. Removing the weakest server usually saves more context than any tuning does.
A good Claude Code setup is small: Context7 for docs, GitHub for code, Playwright for the browser, and one server for your data or your errors. Add a visual server when a project needs one, and leave the rest of the market alone.
If you want the visuals that go with that setup, try PicassoIA Image for a first generation and PicassoIA Image Editor Pro to adjust it. Write one prompt for a hero image, change the lighting or the lens, and compare. The second attempt is where most people find a style they like. Browse every model at picassoia.com/en/all-models and start your first image today.
