Field notes

Notes on what we observe when we check MCP servers, and the fixes for what we find.

See what we observed on the servers we track →

2026-10-27MCP error -32001: Request timed out, -32000: Connection closed, and other errors your client makes up"MCP error -32001: Request timed out" and "MCP error -32000: Connection closed" come from your MCP client's SDK, not from the server. What each one means, how to raise the 60-second timeout, and what changes in v2 of the TypeScript SDK.

troubleshooting protocol

2026-10-27MCP error codes explained: protocol errors vs tool errorsWhich JSON-RPC and MCP error codes a server should send, when a failed tool call belongs in an isError result instead, and what reaches the model in each case.

protocol troubleshooting

2026-10-25An unknown tool should get a JSON-RPC error, not isErrorA maintainer fixed the one observed risk on their MCP Checkup report. What the error_taxonomy check sends and expects, how to test your server with curl, and how often 3,309 servers answered an unknown tool with isError.

maintainers protocol troubleshooting

2026-10-23Use MCP Checkup inside Claude before you connect an MCP serverMCP Checkup is now in Claude's connector directory. How to add it, what to ask it about an MCP server before you connect one, what comes back, and what it does not do.

directory attestations

2026-10-21What 119,510 signed MCP server results show about uptime and tool-list driftWhat our probe saw across 5,916 public MCP endpoints in the month to 21 October 2026: how often they answered, how often their tool lists changed, and which protocol revisions they use.

data attestations

2026-10-21MCP server observations, October 20265,916 remote MCP endpoints on 21 October 2026: 60% signed in full, 32% signed as auth-gated, 8% unsigned; protocol revisions and 1,487 tool-list changes.

data attestations

2026-10-20Streamable HTTP vs SSE: moving a remote MCP server to the current transportHow HTTP+SSE (2024-11-05) differs from Streamable HTTP, what 2026-07-28 changed again, how to keep old clients working, and curl checks for the move.

protocol maintainers

2026-10-17From one loop to nineteen: we measured Cloudflare Queues before running our checks on itOur probe checked servers one at a time. We measured how queue consumers ramp up, fitted a model, tested its predictions, then went to 19 messages a round.

engineering scheduling

2026-10-16MCP server returns 406 Not Acceptable: what it means and how to fix itFix "Not Acceptable: Client must accept both application/json and text/event-stream" (HTTP 406) from an MCP server: the exact SDK messages, how to reproduce them with curl, and the fix in the client, the server and the proxy.

troubleshooting protocol

2026-10-16How many tools does a remote MCP server expose? The median is sevenOf the 1,903 remote MCP servers whose tool list we could read, the median exposes 7 tools, 40% expose five or fewer, and 28 expose 100 or more.

data protocol

2026-10-15The round the platform cancelled and ran again, and 52 minutes of HTTP 503 on /statusA cancelled probe round leaves a job that never finishes, and /status reported it as stuck for 52 minutes. What counts as stuck now, and what we kept.

engineering scheduling

2026-10-14How often does an MCP server change its tools? One in five did within two weeksWe read 1,893 remote MCP servers' tool lists every 12 hours. 22% changed at least once; most changes only added tools, and a few servers made most of them.

data maintainers

2026-10-13The remote MCP server launch checklistThirteen checks before a remote MCP server goes live, from TLS to a toolset fingerprint per release, each with why it matters and a command to verify it.

maintainers protocol

2026-10-13Tool descriptions that give orders to the agent reading themAn AI agent reads every MCP tool description before it picks a tool. Our probe flags a few patterns in that text; 409 of 26,575 tools matched one.

data protocol

2026-10-12MCP server won't connect? Work through the error you getYour MCP server won't connect? Work through the error you get, from 401, 404 and redirects to timeouts and "MCP error -32000: Connection closed", with the curl command for each step.

troubleshooting maintainers

2026-10-11Seven thousand tool snapshots, copied to cold storage byte for byteWe copied 7,102 tool snapshots from our database into gzip objects in cold storage. How we checked the copies, and what has not been copied yet.

engineering data

2026-10-10How to test an MCP server before you ship itA layer-by-layer way to test a remote MCP server from the outside — transport, handshake, tools/list, errors, auth and tool descriptions — with curl commands you can copy.

maintainers protocol troubleshooting

2026-10-10A bug that only appeared in parallel: four servers checked twice, a fraction of a second apartOur first parallel probe rounds checked four servers twice, 0.11 to 0.24 seconds apart. Why one loop never hit it, why we rolled back, and the fix.

engineering scheduling

2026-10-09The words our blog will not store, and why this post cannot name themSince 2026-10-08 new posts on our blog live in a database whose trigger refuses to publish any post holding a verdict word, this one included.

engineering methodology

2026-10-04How to test your MCP server before you ship itPaste your endpoint into the MCP Checkup homepage and see what an agent would see: handshake, tools/list, error shape, redirects, tool text. What each check can conclude, why the 24-hour result is unsigned, and why baseline checks show UNVERIFIED.

methodology trial

2026-10-01Your agent added an MCP server by itself. What do you actually know about it?Personal agents now add MCP servers on their own. No one can see from the outside what a server does with what your agent sends it, but some things can be checked: whether it answers, which tools it offers, and whether they changed. 236 of 1,470 servers in our directory did in 30 days.

directory data methodology

2026-09-30Which MCP protocol revision do servers answer with?Every probe starts by agreeing on a protocol revision, and every signed result records the one it got. Across the 1,004 servers whose latest result is signed, 580 ended on 2025-06-18, the revision our fallback offers, and 394 on 2026-07-28. What those two numbers do and do not say.

protocol data methodology

2026-09-30Five hours when our probe ran nothing, and nothing told usOn 2026-09-23 the scheduler that starts our probe every five minutes stopped calling it. An unrelated deploy brought it back five hours later, and we only noticed nearly four hours after that, reading hourly counts. What happened, what we can and cannot say about why, and the two signals we added.

scheduling engineering

2026-09-30Three in ten public MCP endpoints ask for a key before anything elseOf the 2,282 public MCP endpoints we observe, 694 answer our anonymous handshake with an authentication challenge. We never send credentials, so we have never seen their tools. What we can still observe about them without a key, and what that observation cannot tell you.

directory protocol data

2026-09-30The one call our probe sometimes holds backTo see how an MCP server refuses an unknown tool, our probe calls a name reserved for a tool that should not exist. A server could define that name, so the probe checks the tool list first and holds the call back when it cannot rule that out: 838 of 2,282 public endpoints on 2026-09-30.

methodology protocol data

2026-09-30A rate that could only ever say 100 percentOn 2026-09-23, all 833 servers in our directory showed the same reachability figure: 100 percent. Not because they had all answered us every day, but because of how we counted. We found it while fact-checking a blog post. What was wrong, what we changed, and what the figure shows now.

methodology data engineering

2026-09-24The null byte that would not storeOne tracked MCP server published a legal regular expression containing \u0000 in a tool schema. PostgreSQL cannot hold that character, so for four days every snapshot of that server failed after its header row was written. How we fixed it without rewriting the evidence.

engineering methodology

2026-09-24The probe round that was cancelled at 67 secondsThree probe rounds died halfway through, and the servers left behind were not re-checked for half an hour. The cancellation was the platform doing something ordinary. The half hour was our own guard. How the window went from 30 minutes to 15, and why not to 10.

engineering scheduling

2026-09-24Which code checked you: suite_commit and suite_digestEvery signed result names the check-suite commit that produced it and a digest of that commit's check-suite tree, which anyone with the public repository can recompute with one command. How the value is made, how the probe is stopped from deploying a stale one, and what older results cannot offer.

attestations methodology

2026-09-24Sign-in was broken for two days and nothing logged itFrom 12 to 14 September neither Google sign-in nor the email link worked on mcpcheckup.com. Every server-side log was clean. The cause was a build with two empty variables, made from a clean checkout on purpose. What the fix was, and the gate that now refuses to deploy that build again.

engineering

2026-09-24Three ways our first chart was wrong before it shippedThe directory statistics page has one small line chart: signed results per day, last 90 days. In review it turned out to cover the last 90 rows, not days; its timestamps could be read in the wrong time zone; and on a wide screen the line did not sit over its labels. All three fixed before merge.

engineering

2026-09-24We asked 1,144 MCP servers for a made-up toolThe one tool call our probe makes names a made-up tool, to see how the server says no. Until this week we recorded 58% of signed servers as a risk for that check and could not say what they had answered. Now we record seven kinds of answer, and nine in ten of the flagged servers gave the same one.

methodology protocol data

2026-09-23Your MCP server may already be in a directory you never signed up forMCP Checkup lists remote MCP servers it observes, whether or not anyone registered them with it. How a server gets in, what the page says about it and refuses to say, how to take ownership of the page, and how to opt out.

directory maintainers

2026-09-23Why we don't score MCP serversMCP Checkup shows what it observed about an MCP server and refuses to roll it into a number or a grade. The reasons are structural, not modesty.

methodology maintainers

2026-09-23Why we probe unclaimed servers every 12 hoursMCP Checkup checks unclaimed servers twice a day, claimed servers on the Free plan every 8 hours, and Indie servers every 15 minutes. The reasoning behind each number, what the cadence does and does not tell you, and what money does and does not change.

methodology scheduling

2026-09-23A test that failed two percent of the timeAn open-redirect guard in our test suite went red at random, about one run in fifty. The cause was a substring check against a 43-character token. The first fix would have made the guard blind. The lesson is about what a guard's comment owes the reader.

engineering

2026-09-23The spinner that never stopped: back/forward cache and a waiting state with no way homeSubmit the homepage check, press Back, and the button came back disabled, its label frozen mid-count, with nothing able to re-enable it. The cause was the browser's back/forward cache restoring a page mid-wait. The fix: a small controller with a pageshow reset, and removing a beforeunload listener.

engineering

2026-09-23Our CI builds the site with none of its keys, on purposeType checks and unit tests were green while a stray test file in the pages directory would have broken the production build. So CI now runs a real astro build, and injects no secrets to do it, because no prerendered route is allowed to read the database at build time.

engineering

2026-09-14Claiming an MCP server page: we verify the address, not the recordHow MCP Checkup decides who may speak for a server page. Proof is anchored to the endpoint host itself, via a well-known file or a DNS TXT record, so a claim can never be transferred between records or faked with a repository you happen to control.

maintainers claiming

2026-09-14What a signed attestation proves, and what it does notA probe run on MCP Checkup produces a signed DSSE envelope unless it is disqualified, and the page says so. Here is exactly what the signature covers, how the code that ran is identified, how a third party can check it, and the honest list of things the signature says nothing about.

methodology attestations

2026-09-14The homepage check leaves no trace in the directoryPaste an MCP endpoint on the MCP Checkup homepage and you get a one-off result page that expires. What that check does, what it deliberately does not do, and why a trial is never allowed to become a public page on its own.

methodology trial

2026-09-14Why MCP Checkup doesn't run your code (and what that means for stdio servers)A remote MCP server is an address. A stdio server is a package, and checking it means executing it on somebody's machine. Where the line sits today, and the three-tier design we are building toward.

methodology stdio

2026-09-14The several ways an MCP server says "no", and which ones we count against itA refusal is not a failure. MCP Checkup distinguishes a well-formed JSON-RPC error, a credential gate with a real challenge, a modern protocol rejection, and a malformed refusal, and records each differently. Here is the taxonomy the probe uses.

methodology protocol

2026-09-14Every production migration now runs twice. The first time, we roll it back.How MCP Checkup applies schema changes to a live Postgres: a full dry run inside one transaction that ends with a deliberate exception, then the real apply. What the dry run measures, and the two things it caught this month.

engineering

2026-09-01Why does my MCP server show "observed-unsigned"?Five real protocol-level causes behind an "observed-unsigned" MCP Checkup report — four in a server's own setup, one a bug in our probe that we found and fixed. Each with a curl self-test and a fix.

troubleshooting faq