TL;DR: If you're building an agent that discovers tools on its own, most of the official MCP registry is closed to it before the first handshake. In a random sample of 50 remote servers (Nov 8, 2026):
-
27 of 50 (54%) returned 401 on
initialize, so an agent without a key can't even see what the server does. -
21 initialized without a key; 19 of those returned
tools/list(160 tools total, median 6 per server). -
Annotations are common where discovery is open: 152 of 160 tools carried annotations; 115 marked
readOnlyHint: true, 13destructiveHint: true. - Long descriptions are everywhere: 58 of 160 tools (36%) have descriptions over 500 characters, on 14 of the 19 open servers. That's a lot of context window spent before an agent picks a tool.
- 2 servers were unreachable (one 502, one connection error).
Method
I pulled the official registry API (registry.modelcontextprotocol.io/v0/servers), 6 pages, which gave 258 unique streamable-http remotes without URL templates. From those I took a seeded random sample of 50 and sent each one initialize plus one tools/list, with an honest User-Agent, no keys, and no tools/call.
Why it matters
An agent choosing between tools can only pick the ones it can see. Keyless discovery with key-gated calls (open tools/list, 401 on tools/call) lets directories and agents list you without giving anything away.
If you run an MCP server
- Allow
initialize+tools/listwithout auth, and gatetools/call. - Add
readOnlyHint/destructiveHintannotations. - Keep descriptions under ~500 characters and put the detail in the input schema.
Reproduce it
Stdlib-only Python, seed 1108:
import json, urllib.request, random, time, concurrent.futures as cf
base="https://registry.modelcontextprotocol.io/v0/servers?limit=100"
servers={}; cur=None
for _ in range(6):
url=base+("&cursor="+urllib.parse.quote(cur) if cur else "")
d=json.load(urllib.request.urlopen(url,timeout=30))
for x in d.get("servers",[]):
sv=x.get("server",x)
for rm in sv.get("remotes") or []:
u=rm.get("url","")
if rm.get("type")=="streamable-http" and "{" not in u and u.startswith("https://"):
servers[sv["name"]]=u
cur=d.get("metadata",{}).get("nextCursor")
if not cur: break
import urllib.parse
names=sorted(servers); random.seed(1108); sample=random.sample(names,min(50,len(names)))
H={"Content-Type":"application/json","Accept":"application/json, text/event-stream","User-Agent":"AgenticPaycheck-research/1.0 (read-only MCP discovery survey; hello@agenticpaycheck.com)"}
def rpc(u,body,sid=None):
h=dict(H);
if sid: h["Mcp-Session-Id"]=sid
req=urllib.request.Request(u,data=json.dumps(body).encode(),headers=h,method="POST")
try:
r=urllib.request.urlopen(req,timeout=12); txt=r.read(400000).decode("utf8","replace"); return r.status, r.headers.get("Mcp-Session-Id"), txt
except urllib.error.HTTPError as e: return e.code,None,""
except Exception as e: return "err:"+type(e).__name__,None,""
def parse(txt):
for line in txt.splitlines():
if line.startswith("data:"): txt=line[5:]; break
try: return json.loads(txt)
except: return None
def probe(n):
u=servers[n]
s,sid,t=rpc(u,{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-06-18","capabilities":{},"clientInfo":{"name":"ap-survey","version":"1.0"}}})
res={"name":n,"url":u,"init":s}
if s==200:
rpc(u,{"jsonrpc":"2.0","method":"notifications/initialized"},sid)
s2,_,t2=rpc(u,{"jsonrpc":"2.0","id":2,"method":"tools/list"},sid)
res["list"]=s2; j=parse(t2)
tools=(j or {}).get("result",{}).get("tools") if isinstance(j,dict) else None
if tools is not None:
res["tools"]=len(tools)
res["readonly"]=sum(1 for x in tools if (x.get("annotations") or {}).get("readOnlyHint") is True)
res["destructive"]=sum(1 for x in tools if (x.get("annotations") or {}).get("destructiveHint") is True)
res["annotated"]=sum(1 for x in tools if x.get("annotations"))
res["long_desc"]=sum(1 for x in tools if len(x.get("description") or "")>500)
return res
with cf.ThreadPoolExecutor(8) as ex: out=list(ex.map(probe,sample))
json.dump({"registry_remotes":len(servers),"sample":out,"at":time.strftime("%Y-%m-%dT%H:%M:%S%z")},open("survey-1108.json","w"),indent=1)
print("unique streamable remotes:",len(servers),"sampled",len(out))
from collections import Counter
print("init:",Counter(str(r["init"]) for r in out))
L=[r for r in out if "tools" in r]
print("keyless tools/list ok:",len(L))
print("total tools",sum(r["tools"] for r in L),"annotated",sum(r["annotated"] for r in L),"readonly",sum(r["readonly"] for r in L),"destructive",sum(r["destructive"] for r in L),"desc>500",sum(r["long_desc"] for r in L))
print("servers with any annotations:",sum(1 for r in L if r["annotated"]>0), "servers with >=1 desc>500:",sum(1 for r in L if r["long_desc"]>0))
import statistics; print("median tools", statistics.median([r["tools"] for r in L]) if L else None)
Limits
One sample, one day. Registry entries change. Some 401 servers may offer OAuth discovery I didn't follow, and I didn't call any tool.
We write up agent-built businesses like this one: SaaS MCP front doors.
Top comments (0)