Inside the TokensCOST MCP Servers Directory: 116+ Vetted Servers, Ranked and Searchable

The MCP Ecosystem Got Big — Fast

A year ago you could fit every production-grade Model Context Protocol server on a single screenshot. Today there are hundreds of public MCP servers across thirteen categories, and the official Anthropic registry plus the half-dozen community indexes all give you the same problem: a wall of names, no consistent metadata, and no way to tell a 200-star toy from a 30k-star tool that ships in Cursor by default.

That's why we built the TokensCOST MCP Servers Directory — a single, opinionated index of every MCP server worth knowing about, with the metadata developers actually need to make a 60-second decision.

At launch it covers 116+ verified servers across Developer Tools, Web & Search, AI & ML, Productivity, Databases, DevOps, Communication, Files & Documents, Design, Data & Analytics, Finance, Entertainment, and Other. Every entry is reviewed, categorized, and linked to its canonical repo.

What the Directory Gives You That a Raw Repo List Doesn't

The whole point of an MCP server is to plug a capability — search, a database, a SaaS API, a filesystem — into an LLM agent. Picking the wrong one costs you an afternoon of failed installs and, if it actually runs, a quietly inflated token bill from over-eager tool calls.

We index six things for every server so you can rule one in or out without leaving the page:

1. Live GitHub stars — pulled fresh, not frozen at submission time. Sort by popularity and the top of the list always reflects what the community is actually using.

2. Last-pushed date — the single best abandonment signal. A server with 8k stars and no commits in 14 months is a different bet than 800 stars updated yesterday.

3. Category + verified badge — hand-classified, so "Web & Search" is actually web and search, not "anything that calls an HTTP endpoint."

4. Tools and resources exposed — the named methods the server publishes over MCP, with one-line descriptions. This is the field most registries skip and it's the one that tells you whether the server actually does what its README claims.

5. Install snippet — the exact `npx` / `uvx` / Docker command, ready to paste into your client config.

6. Source link to the original listing — every entry credits and links to its upstream registry so you can audit our edits.

All of it is searchable, filterable by category, sortable by stars / name / category / recency, and paginated with infinite scroll. Works on mobile. No login. No paywall.

Why This Matters for Token Costs

A blog about token pricing covering an MCP directory isn't a stretch — the choice of MCP server is one of the biggest hidden levers on agent cost, and almost nobody talks about it that way.

Three concrete patterns we keep seeing:

1. Chatty servers blow up tool-call budgets

An MCP server that returns 8 KB of JSON per call versus one that returns 800 bytes is a 10× multiplier on the *input* tokens your model sees on the very next turn. Multiply by 30 tool calls in an agent loop and you've doubled the bill for the same task. The directory surfaces the tool list so you can tell at a glance whether a server is going to flood your context window — see The Context Window Cost Trap for the math.

2. The wrong search backend triples your retrieval cost

There are now five mainstream "web search" MCPs (Brave, Tavily, Exa, Serper, Perplexity-style). Their per-query prices vary by ~6× and their result-payload sizes vary by ~4×. Pick the wrong one and your RAG agent quietly costs 3× more per question with no quality difference. The directory lets you compare them side-by-side in one tab.

3. Unmaintained servers force fallback to flagship models

A flaky or abandoned MCP server makes agents retry, escalate, or fall back to the heaviest model in the cascade — the exact failure mode that destroys the cascade-routing savings we covered in Save Money on AI Models with Smart Settings + Routing. The "last pushed" column is the cheapest possible defense against this.

Who the Directory Is For

We built it for three audiences specifically:

• Engineers adding their first MCP servers to Cursor, Claude Desktop, or Cline. Stop scrolling READMEs at midnight. Filter by category, sort by stars, copy the install line.
• Platform teams standardizing internal agent tooling. Use the directory as a curated allow-list. Every server has a stable canonical link and a verified badge.
• Anyone evaluating MCP itself. Browsing the directory is the fastest way to grasp the *shape* of the ecosystem — what categories are deep (Developer Tools, Web & Search), what's still thin (Finance, Design), and where the real momentum is.

What's Next

A few things we're shipping in the next two weeks:

• Per-server token-cost notes — average tokens returned per tool call, measured on a fixed test harness, surfaced right next to the stars.
• "Used by" badges — which clients (Cursor, Claude Desktop, Cline, Continue) ship the server in their default config.
• Submission flow — open an issue / PR to add a server. Every new entry will go through the same metadata review.

If a server is missing or out of date, ping us — the index is meant to be a living document, not a snapshot.

Try It

The directory lives at /mcp-servers. Bookmark it, share it, and if you find one server that saves you an afternoon, the page paid for itself.

Related Reading

• Cost-Efficient AI Agents: Settings & Orchestration — Per-role token budgets and tool-call discipline.
• The Context Window Cost Trap — Why MCP-server payload size matters more than you think.
• Save Money on AI Models with Smart Settings + Routing — The cascade pattern unmaintained tools quietly break.
• Agent Loop Cost Estimator — Plug your tool-call counts in and see the bill before you ship.