iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
In Pennyforge Studio’s October 2026 probe, 40 of 78 MCP endpoints that answered initialization—51.3%—went on to complete an anonymous sequence of initialization, tool discovery, and one safe, read-style tool call. That is a result for a small, selected group of public endpoints, not a claim that half of all MCP servers fail. A separate study of 400 sampled npm packages found that 195, or 48.8%, completed a one-shot startup handshake; it measured a different population and stopped before calling a tool.
What does it mean for an MCP server to work?
A server can respond to an initial request yet still fail at a later step an integration needs. The Model Context Protocol (MCP) specification, versioned 2025-11-25, describes MCP as an open protocol for connecting LLM applications to external data sources and tools. It uses JSON-RPC 2.0 messages among hosts, clients, and servers, with capability negotiation and server-provided tools, resources, and prompts.
That makes “works” dependent on the test. A handshake checks whether the client and server can establish the protocol session and discover tools. A functional tool-use probe goes further and attempts an invocation. Neither check, on its own, establishes long-term availability, production reliability, or security.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →What Pennyforge’s 51.3% result actually measured
Pennyforge Studio reported its probe on October 6, 2026. On October 2, it examined 186 endpoints in an alphabetical a–b slice of one public registry. Of those, 87 returned HTTP 401; the others errored or timed out before inclusion in the measured cohort. The report’s cohort contained 78 endpoints that answered an initialization call.
#1 Best Overall
- More for the money with this high quality Product
- Offers premium quality at outstanding saving
- Excellent product
- 100% satisfaction
For each endpoint, the probe attempted an anonymous path: initialize, then tools/list, then one safe call. It selected the first declared tool with no required input properties and a read-style name, such as get, list, or search. Forty endpoints completed that full path: 40 of 78, or 51.3%.
The reported outcomes across the 78 were:
- 40 completed the full anonymous path.
- 19 returned tool-call errors.
- 15 returned a tool list, but no successful tool call was reported.
- 2 had tool-list errors.
- 2 had initialization errors in the reported outcome breakdown.
The report describes 12-second caps, eight parallel workers, and a test lasting about 20 minutes. Its categories measure the result of that bounded probe, not the condition of every tool on a server. In particular, the probe does not establish whether an endpoint that requires credentials would work after valid authentication, and it says nothing about longer-term uptime or production behavior.
Why another study also found roughly half—but tested something else
A September 2026 paper by Haseeb Mohammed Afsar is summarized in an October 2026 article by Masato Okuwaki of CORe Inc. The study used a different frame: a snapshot of 24,135 registry servers dated August 22, 2026, from which 7,258 active npm packages declaring stdio transport formed the candidate population. It drew 400 packages without replacement using a published seed and frame hash.
Free tools Windows power users keep installed
One-click scans. No signup required.
Each package was launched once over stdio with no credentials and no retry. The probe initialized, listed tools, and validated schemas, but did not invoke a tool. The account reports 195 successful handshakes (48.8%), 150 failures (37.5%), 53 packages requiring credentials (13.3%), and 2 unavailable packages (0.5%). Because a transient timing or network problem still counted as a failure, the paper’s reported 48.8% is a floor for the share that might start, rather than an estimate of enduring availability.
Rank #3
- Product type: Screw kit
- Made by Super Micro
- Manufacturer part number: MCP-410-00005-0N
- Supermicro MCP-410-00005-0N Screw Bag(100PCS) and Label for 24x Hot swap
- Mfr Part Number: MCP-410-00005-0N
The article also reports 66.7% handshake completion in a separate 24-server curated comparison frame under the same probe. That contrast illustrates how a curated list can produce a higher first-attempt startup result; it does not show that the curated servers are more reliable in production. The npm/stdio frame represented 30.7% of the reported registry population. The study did not test remote servers, PyPI packages, or OCI images, and its registry metadata was self-declared.
How to compare the two “about half” figures
| Probe | Population and denominator | Success criterion | Credentials and retries |
|---|---|---|---|
| Pennyforge Studio, 2026 | 78 endpoints that answered initialization from one public registry’s a–b slice | Initialization, tool listing, and one anonymous safe read-style call; 40 of 78 completed (51.3%) | No credentials; retry policy not stated in the report |
| Afsar study, as summarized by CORe Inc., 2026 | 400 sampled active npm packages declaring stdio transport | One launch, initialization, tool listing, and schema validation; no tool invocation; 195 of 400 completed (48.8%) | No credentials and no retry |
The percentages are not interchangeable: they have different sampling frames, denominators, and definitions of success. One measures completion of an anonymous tool call after preselecting endpoints that answered initialization; the other measures a one-shot handshake across sampled npm/stdio packages.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What the findings mean for an integration team
- A registry entry is not an end-to-end test. Discovery or an initial response does not confirm that the tools your client needs can be listed, validated, and invoked.
- Plan for startup diagnosis as well as authentication. In the 400-package study, more sampled packages failed the handshake (150) than required credentials (53). Those counts describe that sample and method, not every MCP deployment.
- Keep readiness and security checks separate. A successful handshake is not a security audit. The CORe article says the paper makes no security claim. It also reports that 1,626 of 2,766 tools lacked safety annotations; absent annotations are missing information, not evidence that a tool is safe or unsafe.
A practical way to check a configured server
A useful readiness check should mirror the steps your application depends on. The CORe article recommends scheduled handshake probes in CI; that is a recommendation, not evidence that one schedule or probe design guarantees production reliability.
- Launch or connect. Use the same transport and configuration as the integration, then record whether the process or endpoint becomes reachable and how long it takes.
- Initialize. Confirm the protocol session completes and record the negotiated capabilities. Separate a failed launch from an initialization failure.
- List and validate tools. Record the returned tools and check their declared schemas before relying on them. A successful list alone does not prove a tool can run.
- Test authentication as a distinct case. If normal use requires credentials, run an authorized check with the intended credentials. An anonymous failure or HTTP 401 does not establish that the authenticated path is broken.
- Invoke only when safe and authorized. Where the tool’s behavior is understood, test a read-only operation with valid inputs. Do not infer safety from a tool name or missing annotations.
- Repeat when measuring availability. A one-shot check can count a transient failure as a failure. Keep initial-attempt results distinct from retry or scheduled-monitoring results, and retain the timestamps and error categories.
These checks answer different operational questions: whether a server starts, whether it exposes usable tool definitions, whether the intended credentials work, and whether a permitted operation succeeds. Reporting those stages separately makes a failure easier to diagnose than a single pass/fail label.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
