Your Claude Code session is a few hours in, and the MCP tools just... stop existing. No error, no crash, no red text. Claude tries to use a tool from your database server or your filesystem server and it's simply not there anymore. If you check the MCP client log, you'll find this pair of lines from a few hours back:
Sending SIGINT to MCP server process
MCP server process exited cleanly
That's the whole crime scene. If you're on Claude Code 2.1.207 and the server was stdio transport, there's an open issue that describes exactly this — with better evidence than most bug reports I've read.
What's actually reported
#76769, filed 11 July, labeled regression and has repro on the tracker. One user, running long-lived sessions on Linux, found that starting in 2.1.207 every stdio MCP server gets a clean SIGINT at spawn + 4 hours — and is never respawned. The session keeps going. The tools are just gone.
Three details make this report worth taking seriously:
It's a connection-age cutoff, not an idle timeout. Across three independent sessions, the kill landed at spawn + 4h00m, give or take seconds — whether the session was busy or idle at the time. So "keep the session active" doesn't save you.
There's a clean version bisect. The reporter read each session's exact binary version from /proc/<pid>/exe rather than guessing. On 2.1.202 a server stayed alive 90+ hours. On 2.1.206, 29 hours and counting. On 2.1.207, dead at 4h00m, every time. That's about as tidy as a community bisect gets.
The kill is permanent for the session. Claude Code doesn't auto-reconnect stdio MCP servers after a disconnect — that's long-standing behavior Anthropic closed as not-planned in #43177. So once the 4-hour SIGINT lands, the only way to get your tools back is a full session restart. Which re-arms the same timer.
To be straight about confidence: this is one user's report and it isn't Anthropic-confirmed — no maintainer had responded as of 12 July. But I checked the 2.1.207 changelog and there's no mention of any MCP lifetime change, so if it's intentional, it's undocumented. I also checked npm: 2.1.207 is the latest release as of today, so there's no fixed version to upgrade to yet.
Why this one is nasty
The failure is silent. If you run unattended or overnight sessions that lean on MCP — a database server, AWS tooling, filesystem access — the session doesn't stop when the server dies. It keeps working without those tools, and depending on what you asked for, it may just... route around them. You find out hours later, looking at output that quietly never touched your database.
I run long automation sessions, which is exactly the profile this hits. Four hours is not a long session when Claude is babysitting a build or monitoring something.
What to do about it
- Check if you're exposed. Run
claude --version. If you're on 2.1.207 and use stdio MCP servers in long sessions, you're in the blast radius. Mid-session,/mcpshows whether your servers are still connected. - Restart before the 4-hour mark. It's the only recovery in the report. Ugly, but it works — just remember the timer re-arms.
- Pin to 2.1.206 if you installed via npm:
npm install -g @anthropic-ai/claude-code@2.1.206, and setDISABLE_AUTOUPDATER=1so a background update doesn't undo it. - On a native install, don't assume a downgrade sticks. See below.
- Watch #76769. Claude Code ships several times a week; there's a decent chance the fix lands before your next long session. Check the changelog before taking any of this advice as current.
What not to do
Don't just delete the 2.1.207 binary on a native install. The same reporter tried pinning back to 2.1.206 by removing the binary and repointing the launcher symlink — and found the deleted binary restored itself about 35 minutes later, identical md5, with the symlink pointed back at it. That was with autoUpdates: false set the whole time; they suspect an undocumented native-install self-repair flag. That's one unconfirmed observation, but if you're about to fight the updater manually, know that someone already lost that fight once.
Don't wait for the server to reconnect on its own. It won't. #43177 closed that door as not-planned. If the tools are gone, the session is done having them.
Don't assume every MCP setup is affected. The report covers stdio-transport servers on one user's Linux machines. HTTP-transport servers aren't part of the reported behavior — not because they're confirmed safe, but because nobody's tested them in this issue yet.
Receipts
- #76769 — the regression report: open, labeled
regression+has repro, no Anthropic response as of 12 July - #43177 — stdio MCP servers never auto-reconnect; closed as not-planned
- The 4-hour timing, the version bisect, and the self-restoring binary are all one user's analysis — solid methodology, but not Anthropic-confirmed. I verified the issue contents, the changelog, and the current release version myself on 12 July against 2.1.207.
If your MCP tools vanish mid-session, now you know where the four hours went. Check your version before your next overnight run.
More fixes like this, before they hit the blog