Essential MCP trio — the 3 servers that cover most workflows

Scorecard question: Which MCP servers do you actually use in a typical week? (check all that apply — 1 pt each, max 3 pts) Max‑score answer (3 pts): GitHub MCP + Context7 + Figma MCP (or swap Figma for Sentry / Linear if you don’t implement designs).

Why this matters in 2026

By May 2026 the MCP ecosystem has stopped being a novelty and started being plumbing. The question isn’t “do you use MCP?” anymore — it’s “do you use the right three?” Q9 already established the sweet spot is 3–5 servers (any more and tool-list bloat starts eating your agent’s context budget and degrading selection accuracy). Q10 is about which three. The “essential trio” — GitHub for the repo/PR/issue graph, Context7 for live library docs, and Figma for design-to-code — covers >70% of what a typical IC developer does in a normal week: shipping code against an existing codebase, looking up the API of a library they don’t have memorized, and translating a design into a component. Picking those three (or the right substitution when your work skews different) is the single highest-leverage Q10 move, because it puts MCP power on the three surfaces where most of your turns actually happen. The wrong three — five productivity SaaS connectors and zero code-grounding — looks impressive in your claude mcp list output but does almost nothing for your throughput.

What “max score” actually looks like (the trio for IC developers, plus situational picks)

A max-score Q10 setup is concrete and verifiable. Run claude mcp list (or codex mcp list / Cursor’s MCP panel) and you should see three to five entries — never zero, never fifteen — anchored by GitHub MCP, Context7, and one of {Figma, Sentry, Linear}. For a frontend or full-stack IC who ships designs, the trio is GitHub + Context7 + Figma. For a backend or platform engineer who lives in production observability, swap Figma for Sentry. For a developer who works mostly off tickets and rarely touches the Figma file directly, swap Figma for Linear (or Atlassian/Jira). The trio scores 3 pts because each server unlocks a category of work that the agent otherwise has to guess at: GitHub gives it the real PR/issue graph instead of a fuzzy memory of what’s in git log, Context7 gives it the current docs for the library you’re using instead of last-year’s API surface from its training data, and Figma (or its substitute) gives it the actual design spec / error stack / ticket text instead of a screenshot. A scorecard taker who installed six MCP servers but none of them are these three scores worse on Q10 than one who installed exactly these three, because the trio is selected to cover the categories developers actually spend time in, not the categories that look good in a screenshot.

Current landscape (web‑search‑verified)

By mid-2026 the MCP marketplace has thousands of entries, but the day-one configuration that consistently shows up in “best MCP servers” roundups is GitHub + Context7 + a third pick that varies by role. The signal isn’t that these three are the best on any single axis — it’s that they’re the three with the broadest coverage, where adding a fourth gives diminishing returns until you hit a project-specific need.

GitHub MCP — repos, PRs, issues

The single highest-impact install for almost every developer. GitHub MCP turns the agent into a participant in your actual issue/PR workflow: it can search code across your orgs, read PRs and reviews, open issues, comment on PRs, and (with the right permissions) push branches. The official remote endpoint is hosted by GitHub Copilot’s infrastructure, so you skip the local-server setup entirely. Install it with the HTTP transport — this is the canonical 2026 install command:

claude mcp add --transport http github https://api.githubcopilot.com/mcp/

After install, run claude mcp list and confirm github is listed. Then in a session, ask the agent something like “find the open PRs in this repo where I’m a reviewer and summarise the requested changes” — if it works, you’re done. The non-obvious win here is not the obvious one (asking the agent to open PRs for you) but the grounding effect: when the agent has GitHub MCP, it stops hallucinating about what’s in your codebase and starts reading the real thing.

Context7 — up-to-date library docs

Context7 is a remote MCP server that fetches live, version-specific documentation for thousands of libraries and frameworks and injects it into the agent’s context on demand. It’s the fix for the “the agent’s training data is six months stale” problem: instead of guessing at the current useEffect signature or the latest Tailwind utility name, the agent calls Context7 and pulls the current docs. Free tier covers most usage; paid tiers raise rate limits. Install with the HTTP transport:

claude mcp add --transport http context7 https://mcp.context7.com/mcp

The killer pattern is appending “use context7” to a prompt where you’re working in a library you don’t have memorized — the agent picks up the cue, pulls the relevant docs, and writes code against the real current API instead of a plausible-looking hallucination. Context7’s biggest payoff is on fast-moving stacks (React 19, Next.js 15+, Tailwind 4, Vite 6, anything in the Cloudflare ecosystem) where the model’s training cutoff is most likely to be wrong about the current API.

Figma MCP — design handoff (only if you implement designs)

The official Figma Dev Mode MCP server exposes the live structure of whatever you have selected in Figma — hierarchy, auto-layout, variants, text styles, spacing tokens, component references — directly to the agent. Instead of giving the agent a screenshot and hoping it can read pixel measurements, you give it the real spec. This is the highest-leverage MCP for any developer who turns Figma files into React/Vue/Astro components. The server runs locally inside the Figma desktop app, so the setup is a little different from pure remote MCPs: enable Dev Mode MCP in Figma’s preferences, then add the local endpoint:

claude mcp add --transport http figma http://127.0.0.1:3845/mcp

(Port and exact path are published in Figma’s setup docs and may shift — verify against the official Figma “Dev Mode MCP” page before pasting.) Once installed, select a frame in Figma and ask the agent to implement it — the agent reads the auto-layout, the variables, the text styles, and writes a component that matches the design’s actual structure rather than your guess at it.

Situational alternatives

Swap the third slot when your work doesn’t centre on Figma:

Sentry MCP — for backend / platform / SRE work where production errors are the most common context the agent needs. Pulls full issue details, breadcrumbs, environment context, and related events directly from your Sentry org. Endpoint: https://mcp.sentry.dev/mcp (verify the current canonical URL on Sentry’s docs page before installing).
Linear MCP — for teams that live in Linear for tickets and project work. Lets the agent read issue text, comments, sub-issues, and statuses, and update them. Endpoint: https://mcp.linear.app/mcp.
Atlassian (Jira/Confluence) MCP — the Linear-equivalent for the Atlassian stack. Endpoint: https://mcp.atlassian.com/v1/sse was the old form; the current /mcp HTTP endpoint is the one to install — check Atlassian’s docs page for the canonical URL since the legacy /sse form is being deprecated.
Database MCPs — PostgreSQL (@modelcontextprotocol/server-postgres for local Node-based install), or hosted variants like Neon’s and Supabase’s official servers. Worth it when your work is heavily schema-driven.
Cloud provider MCPs — Cloudflare exposes a family of servers under the {name}.mcp.cloudflare.com/mcp pattern (e.g. https://bindings.mcp.cloudflare.com/mcp, https://observability.mcp.cloudflare.com/mcp, https://workers-builds.mcp.cloudflare.com/mcp). Pick one or two per Cloudflare service you actually deploy on, not all of them. Vercel uses just the domain: https://mcp.vercel.com. Stripe similarly: https://mcp.stripe.com.

The rule for the third slot is “pick the one whose category you spend the most time in this week”, not “pick the one that looks most impressive”. A backend developer with Linear in slot 3 because they live in tickets scores the same 3 pts as a frontend developer with Figma in slot 3 — and both score higher than someone with Slack, Notion, and Google Drive in slots 3/4/5.

Step-by-step: installing the trio

Decide your third pick before you start. GitHub and Context7 are non-negotiable. The third is role-dependent: Figma if you implement designs, Sentry if you live in production errors, Linear/Atlassian if you live in tickets, a DB MCP if you work mostly off schema, a cloud MCP if you do deploy/infra work. Don’t install all of them — pick one and earn the 3rd point cleanly.
Install GitHub MCP. Run claude mcp add --transport http github https://api.githubcopilot.com/mcp/. Authenticate when the agent first calls it (you’ll get an OAuth prompt). Confirm with claude mcp list — github should appear.
Install Context7. Run claude mcp add --transport http context7 https://mcp.context7.com/mcp. Free tier is fine to start; sign up for an API key only if you hit rate limits. Confirm with claude mcp list.
Install your third pick. For Figma: enable Dev Mode MCP inside the Figma desktop app preferences, then claude mcp add --transport http figma http://127.0.0.1:3845/mcp. For Sentry: claude mcp add --transport http sentry https://mcp.sentry.dev/mcp. For Linear: claude mcp add --transport http linear https://mcp.linear.app/mcp. Authenticate on first use.
Verify the trio works in a real session. Open a repo, start claude (or codex), and run a multi-MCP prompt: “List my open PRs on this repo (GitHub MCP), look up the current signature for useFormState (Context7), then propose a refactor of the form in src/components/SignUp.tsx.” If all three MCPs fire and the agent’s answer references real PRs + real current docs, you’re done.
Add to .mcp.json for team-wide sharing (optional). If you want every teammate to inherit the trio when they clone the repo, add the configs to .mcp.json at the repo root (committed to git). Local-only servers (like the Figma desktop one) stay in your user-level config; remote ones (GitHub, Context7, Sentry, Linear) can be safely shared via .mcp.json.
Resist installing a fourth. Q9 already established that 3–5 is the sweet spot. If you find yourself adding a fourth, drop one — usually the situational pick that’s no longer matching this quarter’s work. Tool-list bloat is real: every extra MCP adds tokens to every prompt and gives the agent more opportunities to pick the wrong tool.

Common pitfalls

Using wrong endpoints. The /sse path style is deprecated almost everywhere in favour of /mcp HTTP transport (Linear, Sentry, Atlassian, Cloudflare all moved). If you copy-paste a 2024 setup guide, you’ll get /sse URLs that no longer work. Always check the vendor’s current docs page for the canonical endpoint before pasting.
Hallucinated MCP packages. There is no @anthropic/* MCP package on npm — Anthropic does not publish MCP servers under that scope. Likewise @modelcontextprotocol/server-git, server-sqlite, server-filesystem, server-mysql, and server-jira are not on npm (some exist as Python packages installable via uvx, but npm install will fail). The ones that are on npm: @modelcontextprotocol/server-github, server-gitlab, server-slack, server-postgres. If a guide says npm install -g @anthropic/mcp-server-figma, the guide is wrong; the official Figma MCP runs inside the Figma desktop app, not from npm.
Wrong Cloudflare URL shape. Cloudflare uses the {name}.mcp.cloudflare.com/mcp pattern — e.g. https://bindings.mcp.cloudflare.com/mcp, not https://bindings-mcp.mcp.cloudflare.com/sse. The hyphenated-prefix-plus-/sse form is a hallucination that older guides still repeat.
--url instead of --transport http. claude mcp add uses --transport http (or --transport sse for legacy servers) followed by the endpoint URL as a positional argument. There is no --url flag — that’s a flag from an early draft of the docs that never shipped.
Confusing .mcp.json with .claude/mcp.json. The repo-level shared config file is .mcp.json at the repo root. It is not .claude/mcp.json — that path doesn’t exist. User-level config lives under ~/.claude/ but the per-repo file is .mcp.json at the root.
Picking five productivity SaaS connectors and zero code-grounding. A Q10 of “Slack + Notion + Google Drive + Linear + Confluence” looks impressive in claude mcp list but doesn’t ground the agent in your code, your library docs, or your design system — the three categories where it actually saves you turns. If your trio doesn’t include GitHub + Context7, you have a Q10 problem regardless of how many MCPs are installed.

How to verify you’re there

claude mcp list (or codex mcp list) shows 3–5 entries, including github and context7.
The third entry matches your actual day-to-day work: Figma if you ship designs, Sentry if you fight prod errors, Linear/Atlassian if you live in tickets, DB or cloud MCP if your work is schema- or infra-heavy.
A live prompt in your primary repo can fire GitHub MCP + Context7 + the third in a single multi-turn task without you having to babysit which tool gets called.
All endpoints use /mcp HTTP transport (not deprecated /sse), match the vendor’s current docs page, and authenticate cleanly.
No @anthropic/* packages, no *-mcp.mcp.cloudflare.com/sse URLs, no --url flags in your install history.
You can articulate, in one sentence, why each of your three picks is in the trio — and which category it covers (repo graph / library docs / design or errors or tickets).