cd /news/agent-protocols/mcp-servers-that-fit-a-small-context… · home › topics › agent-protocols › article
[ARTICLE · art-170919] src=stork.ai ↗ pub= topic=agent-protocols verified=true sentiment=· neutral

MCP Servers That Fit a Small Context Window: A Token Budget for Local Models

A token-budget analysis of Model Context Protocol (MCP) server tool definitions found that only 39% of MCP servers fit within a 4k-token context window when tool definitions are capped at a quarter of the window, rising to 61% at 8k, 79% at 16k and 90% at 32k, based on a listing of servers as of November 16, 2026. The median server costs about 1,400 tokens across 7 tools, 17% come in under 500 tokens, and 8% exceed 10,000 tokens, putting them out of reach below a 64k window. Ollama's documentation sets default context by video memory at 4k below 24 GiB, 32k from 24 to 48 GiB and 256k at 48 GiB and above, and recommends at least 64,000 tokens for agents and coding tools, while a separate measurement found roughly 57 prompt tokens for a tersely described tool versus about 292 for a verbose one.

read5 min views1 publishedNov 17, 2026
MCP Servers That Fit a Small Context Window: A Token Budget for Local Models
Image: Stork (auto-discovered)

Why this hits local models first# #

When a client connects to an MCP server it loads the definition of every tool, name, description and input schema, into the model's context. On a hosted model with 200,000 tokens you pay for that in cents. On a local model you pay in room, and there is much less of it.

  • Ollama's documentation sets the default context by video memory:4k below 24 GiB , 32k from 24 to 48 GiB, 256k at 48 GiB and above. It recommends at least 64,000 tokens for agents and coding tools.
  • Ollama ships no MCP client of its own, so there is no built-in list of servers to pick from. This comparison of eight clients notes that Playwright's 25 tool definitions alone come to about 3,900 tokens.
  • The cost is paid on every turn, not once. A measurement on Ollama found about 57 prompt tokens for a tersely described tool and about 292 for a verbose one.

Context windows will grow, and this will matter less each year. Today, on the laptop most people actually have, it decides whether tools work at all.

[Read the docs](https://www.stork.ai/developers/talking-head-api?utm_source=house_ad&utm_medium=inline&utm_campaign=api)

[No code? Make a photo talk](https://www.stork.ai/tools/talking-photo?utm_source=house_ad&utm_medium=inline&utm_campaign=api-second)

## The budget[#](#the-budget)

A workable rule is to give tool definitions no more than a quarter of the window. The rest has to hold the system prompt, the conversation, and whatever the tools return.

Share of the MCP servers whose tool definitions we could list on November 16, 2026. One server at a time: connect two and you add their costs.
Context window Budget for tool definitions Share of MCP servers that fit on their own
--- --- ---
4k 1,000 tokens 39%
8k 2,000 tokens 61%
16k 4,000 tokens 79%
32k 8,000 tokens 90%
  • 17% of servers come in under 500 tokens. You can run two of those in a 4k window.
  • The median server is about 1,400 tokens and lists7 tools .
  • 8% are over 10,000 tokens. Those are out of reach below a 64k window.

What well-known servers cost# #

Our own check of each server's public endpoint on November 16, 2026, except Playwright, which runs locally and is quoted from the source linked above. Estimated at 4 characters a token.
Server Tools Tool definitions, estimated tokens Fits a 4k window?
--- --- --- ---
Cloudflare documentation 2 250 Yes
Exa (web search) 2 520 Yes
Context7 (library docs) 2 1,240 Tight
Browser Use 7 1,250 Tight
Playwright 25 about 3,900 (measured by Octomind) No
Firecrawl 3 4,200 No

Two things stand out. The number of tools is a poor guide: Firecrawl lists three tools and costs more than Playwright's twenty-five, because its schemas are long. And many of the best-known servers, GitHub, Linear, Notion, Stripe, Supabase, will not show their tools to anyone who has not signed in, so nobody can tell you their cost in advance. We list them as "sign-in required" rather than guess.

Five ways to make tools fit# #

  • Raise the window if you have the memory. In Ollama, setnum_ctx on the model or theOLLAMA_CONTEXT_LENGTH environment variable. Every extra token of context costs memory, so this has a ceiling.
  • Connect one server per job. Costs add up. Two median servers are already 2,800 tokens.
  • Prefer the lighter server when two do the same job. Among the web search servers we measured, the lightest cost about 250 tokens and others several thousand.
  • Use a client that loads tools on demand , if yours can. Then only the tools in use take room.
  • Check the size before you install. That is the part that used to be impossible, and the reason we measured it.

Enjoying this? Get one like it in your inbox each morning.

one email a day · unsubscribe in two clicks · no third-party tracking

Search by size# #

Every listing in Stork's MCP directory shows the estimated context cost of its tool definitions and the date we last checked it. To search by size from your own agent, connect it to this free MCP server, which needs no key:

https://mcp.stork.ai/mcp?tools=directory

It has five read-only tools and costs about 900 tokens itself. Ask for a server by task and set max_tool_tokens to your budget: 1000 for a 4k window, 2000 for 8k. Each result says how to connect, whether the endpoint answered our last check, what it costs, and what you give up by choosing it, for example a sign-in, or a publisher that is not the vendor. Disconnect it once you have found what you need and the room is yours again.

What we did not measure# #

  • Servers that run on your machine. We check remote endpoints. A local package cannot be measured from outside.
  • Servers behind a sign-in. About a third of endpoints ask for an account before they list anything.
  • What tools return. A server with small definitions can still send back a very large result. Budget for that separately.
  • Whether a tool works. A check is a handshake and a tool listing. It never calls a tool.
The full results of the pass, including how many endpoints answered at all, are in [this report](https://stork.ai/blog/mcp-registry-endpoints-checked-november-2026).

## Frequently asked questions[#](#frequently-asked-questions)

### How many MCP servers can I use with a 4k context?[#](#how-many-mcp-servers-can-i-use-with-a-4k-context)

Usually one, and it has to be a light one. Keeping tool definitions to about 1,000 tokens leaves room for the conversation, and 39% of the servers we measured fit under that on their own. 17% are under 500 tokens, so two of those will fit.

Does Ollama support MCP?#

Ollama serves the model and accepts tool definitions, but it does not include an MCP client. You use a separate client that connects to MCP servers and passes their tools to Ollama.

Why do my local model's tools stop working after I add a server?#

The most common cause is that the tool definitions no longer fit in the context window along with the conversation. Check the size of the server's definitions against your window, raise the window if you have the memory, or switch to a lighter server.

── more in #agent-protocols 4 stories · sorted by recency
── more on @ollama 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/mcp-servers-that-fit…] indexed:0 read:5min 2026-11-17 · —