← Back to Guides · MCP
Best MCP Servers for Claude Code in 2026 (Curated, Not Context-Hungry)
Short answer: "best MCP servers" and "context-hungry MCP servers" are usually the same servers — the problem isn't that any one of them is badly built, it's that every server you connect registers its full tool schema into context on every turn, whether that turn touches it or not. For a pre-PMF solo founder with a small context budget and no second engineer to catch the slowdown, the right shortlist is smaller than most roundups suggest: GitHub, Postgres/Supabase, and Context7 earn a permanent slot; Stripe and Sentry earn one conditionally; everything else — including popular picks like a full Jira/Linear/Slack trio — should usually stay disconnected. This post explains the "context-hungry" test itself, then applies it as a skip list, not just a shortlist.
What counts as "context-hungry," and why it kills small-context workflows
Every MCP server you connect to Claude Code loads its tool definitions — names, descriptions, JSON parameter schemas — into context at the start of a session, before you've typed a prompt. That cost is fixed per server and paid on every turn of the conversation, not just the turns where you actually call one of its tools. A server exposing 20+ granular tools (common in project-management and CRM integrations that mirror a full REST API) can burn a meaningful slice of your context window on schema alone.
That tax matters more, not less, for a solo founder than for a funded team. A well-resourced team can afford a large context window, a fast model, and the latency hit of a crowded tool list, because the volume of work justifies the overhead. Pre-PMF, you're running fast iteration loops — short, frequent Claude Code sessions where you want the model reasoning about your actual code and conversation, not re-parsing a dozen tool schemas for a Jira integration you open twice a month. A context-hungry server isn't "bad" in isolation; it's bad for this specific workflow, where the opportunity cost is your own iteration speed.
The practical test: does this server's tool schema earn its permanent seat in context by getting used most sessions, or does it sit there as dead weight on the 90% of sessions where you don't touch it? If it's the latter, it belongs connected per-project for the week it's needed, or not connected at all — not left on by default "just in case."
The shortlist: servers worth the context budget pre-PMF
| Server | Tool surface | Why it clears the bar | When to connect |
|---|---|---|---|
| GitHub | Small, read-heavy by default | You touch PRs, issues, and CI status every coding day — the highest-frequency use case on this list | Day one |
| Postgres/Supabase | Narrow, mostly read-only | Live schema access catches drift before a migration ships, and there's no second engineer around to catch it in review instead | Day one, read-only role |
| Context7 | Single-purpose docs lookup | Stops the agent guessing at a library API that changed after its training cutoff — low schema cost for what it replaces (a wrong guess you have to debug) | The first time you catch a wrong API guess |
| Stripe | Larger surface, but scoped | Worth it only once billing is live — before that it's schema cost for a tool with nothing real to do | Once billing exists |
| Sentry | Moderate surface | Worth it only once you have production traffic generating real traces to pull into a fix conversation | Once there's production traffic |
These five are covered in more ranking detail, with setup commands, in the ranked MCP servers guide — this post is about the filter you apply before adding anything beyond them, which is the part that roundup doesn't focus on.
Servers to skip (or disconnect) and why
A full project-management suite (Jira, Linear, Asana) connected by default. These tend to expose a large tool surface mirroring their entire REST API — create/update/search/comment/assign across issues, projects, and sprints — because the integration is built for team coordination, not a single person's workflow. As a genuinely solo founder, you're the only one who needs to know your own task state; you already know it. Connect one of these the week you onboard a contractor or co-founder who needs the agent to see shared state, then disconnect it again once that project wraps.
Slack, connected globally. Same shape of problem: a messaging integration's tool schema is sized for "read/send/search across channels and DMs," which is coordination overhead you don't have yet if there's no one else on the other end. Add it per-project when you have a support hire or contractor who needs updates routed there.
A generic "everything" CRM or marketing-automation MCP server. These often bundle contacts, deals, campaigns, and analytics tools into one schema because the vendor built it to cover every seat in a sales org. If you only use two of those tool categories, you're paying full schema cost for the other categories every turn. Prefer a narrower, single-purpose server over a suite server when one exists — or skip it until the workflow is frequent enough to justify the cost.
Any unmaintained or community fork of an official server. This is a different kind of cost — not context, but risk. A stale fork can have an outdated or incorrectly scoped tool schema, and you're trusting its publisher with API credentials. Before connecting anything not published by the vendor itself, run it through the 3-check freshness audit: check the repo's last commit date, open issue backlog, and whether the install command still matches current docs.
Browser automation (Playwright, Chrome DevTools), left on by default. Both are genuinely useful — see the setup commands in the ranked guide — but a backend-only week doesn't need a browser-automation tool schema sitting in context. Connect project-scoped for the week you're shipping and verifying UI changes, then remove it.
The pattern across all of these: it's not that the server is poorly built, it's that its tool surface is sized for a use case — a team, a sales org, a UI-heavy sprint — that isn't your current, default context. Size the servers you keep connected to the workflow you actually run most days.
How to measure your own context cost before deciding
Don't guess at this — check it. The context budget estimator lets you paste your current .mcp.json and see the approximate token cost of your connected servers' schemas before you add or remove anything. Run claude mcp list to see what's actually connected right now (it accumulates faster than most people expect — a server added for one afternoon's task, three months ago, that never got removed), and claude mcp remove <name> to pull one back out. Keep literal API keys out of .mcp.json itself — use ${VAR} references resolved from your shell environment or a local .env kept out of git.
# check what's currently loaded
claude mcp list
# remove anything you added for a task that's done
claude mcp remove <name>
# connect project-scoped (not global) for situational work
claude mcp add playwright -- npx -y @playwright/mcp
How AgenticKit's 34-agent team already wires these in
AgenticKit's agents are written against this same lean-by-default assumption rather than trying to replicate MCP access in code. The backend-architect agent expects live Postgres/Supabase schema access for migration work; the sre-engineer agent expects Sentry for incident triage once there's production traffic; the test-automator agent expects Playwright for end-to-end verification, scoped to the project, during UI weeks. None of that connects automatically — AgenticKit ships the agents, skills, and commands that assume a given server is there once you've connected it yourself with claude mcp add; it doesn't install credentials or servers for you. The kit picker maps this month's priority — backend work, billing, a UI sprint — to the specific agents, skills, and MCP servers that cover it, which is a faster way to decide what to connect than working through a shortlist from scratch each time your priority shifts.
FAQ
Is a "context-hungry" server just a server with lots of tools? Mostly, yes — tool count and parameter complexity both add to the schema size, so a server exposing 20+ granular endpoints costs more per turn than one exposing 3-4 focused ones, independent of how good the underlying integration is.
Should I just connect everything and let Claude figure out what to ignore? No — the cost isn't "the model gets confused," it's that every connected server's schema loads into context on every turn regardless of use, so connecting servers you touch rarely is pure overhead with no corresponding benefit on the sessions where you don't use them.
How is this different from the ranked MCP servers post on this site? That post ranks a core 5-7 servers by daily-use frequency and gives setup commands for each. This post leads with the context-cost filter itself and spends more time on the explicit skip list — what to leave disconnected and why — which the ranked post treats more briefly.
What if I'm not solo — I have one contractor? The moment a second person's work needs to be visible to the agent, a Linear/Slack-style server starts earning its context cost instead of sitting idle; that's the trigger point called out above. Re-evaluate your shortlist whenever your team size or product stage changes, rather than fixing it once.
Does removing a server lose its configuration? No — claude mcp remove <name> only disconnects it from the current session's context; your API keys and .mcp.json entry structure aren't deleted unless you edit the file directly. Reconnecting later is a one-line claude mcp add away.
Next steps
If you've never audited what's connected, start with claude mcp list today — it's common to find a server added for a one-off task months ago still loading its schema on every turn since. Keep the core three (GitHub, Postgres/Supabase, Context7) connected by default, add Stripe and Sentry once their trigger condition (live billing, real production traffic) is met, and treat everything else — project-management suites, Slack, browser automation — as project-scoped additions you remove when the task that needed them is done. Run your current config through the context budget estimator before adding the next one.
Related Tools & Agents
Not sure which servers your stack needs yet?
The kit picker maps this month's priority to the agents, skills, and MCP servers that cover it.
Try the kit picker →