All stdio MCP Servers Fail After hermes update
Bug tracker — July 8, 2026. Hermes bug #61010 exposes a transport-level regression: after a routine update, every stdio-based MCP server fails identically with ClosedResourceError on every tool call. The runtime reports 504 tools available — none are reachable. This is the 7th distinct post-update regression we have tracked in 7 days.
The Failure in One Sentence
After hermes update, every single stdio-based MCP server fails with ClosedResourceError on every tool call. HTTP-based MCP servers continue to work normally. The runtime reports servers as connected (they appear in the tool list with 504 tools available), but cannot communicate with any of them.
Not a Single-Server Problem
Every stdio server fails identically:
| Server | Type | Status After Update |
|---|---|---|
| identity | stdio | ClosedResourceError |
| deepseek | stdio | ClosedResourceError |
| datasheet | stdio | ClosedResourceError |
| pcbparts | stdio | ClosedResourceError |
| RAG | stdio | ClosedResourceError |
| RLM | stdio | ClosedResourceError |
| kicad-mcp-pro | stdio | ClosedResourceError |
| Supabase extensions | HTTP | Working normally |
| Helium Trades | HTTP | Working normally |
The pattern is clear: stdio transport is broken at the framework level. HTTP transport is unaffected.
The Failure Signature
1. Run hermes update 2. Start a session → MCP reload message appears: "MCP servers have been reloaded. Reconnected servers: ... 504 MCP tool(s) now available" 3. Call any tool from any stdio MCP server → ClosedResourceError immediately 4. Restarting the session does not resolve it 5. config.yaml MCP server definitions were not modified
The critical diagnostic detail: ClosedResourceError is raised by the Hermes runtime itself, not by individual MCP server processes. This suggests the stdio transport layer was broken at the framework level by the update — the individual server processes are likely healthy, but Hermes can no longer talk to them.
The Impact
Every user who relies on stdio-based MCP servers — which is most tool-using Hermes deployments — loses all MCP tool functionality after updating. Only HTTP-based MCP servers survive. For a platform whose value proposition includes extensibility through tools, a transport-level regression that silently kills every local tool connection while reporting them as available is a severe reliability failure.
The 7th Post-Update Regression in 7 Days
| Issue | Date | Failure Mode |
|---|---|---|
| #57031 | Jul 2 | Homebrew/Linuxbrew TUI broken — missing ui-tui directory |
| #57790 | Jul 3 | Dashboard web down after hermes update to v0.18.0 |
| #57081 | Jul 2 | Stuck on update for 7 hours on Windows |
| #53069 | Jun 26 | Cannot find module simple-git after upgrade |
| #60929 | Jul 8 | Windows desktop version fails to update |
| #60868 | Jul 8 | Dashboard blank after v0.18.0 update |
| #61010 | Jul 8 | All stdio MCP servers broken after update — transport-level |
Seven distinct update-triggered failures in seven days is not a coincidence. It is a pattern that suggests the update mechanism itself lacks sufficient pre-release integration testing across the full surface area of Hermes functionality.
Why This Matters for Review-Site Readers
- Updates are a reliability event, not a routine operation. When updating your local agent platform carries a meaningful risk of breaking every tool your agent uses, the operational cost of self-hosting becomes materially higher than managed alternatives.
- The failure is silent at the wrong layer. The runtime reports 504 tools available and connected. The tools are listed, visible, and appear healthy. The failure only manifests when you try to use one — and by then you have already built your workflow around tools that do not work.
- HTTP-only survival reveals an architectural fragility. If stdio transport can be broken at the framework level by a routine update while HTTP transport survives, the transport layer lacks the isolation that would prevent one update from taking down an entire class of tool connectivity.
- Seven regressions in seven days. For a comparison shopper evaluating Hermes Agent vs a managed alternative, the update reliability record is a quantifiable operational cost factor.
Case Summary
- Source: GitHub issue #61010
- Opened: Jul 8, 2026
- Labels: bug (no P-label assigned yet)
- OS: Ubuntu 24.04 - WSL2
- Reporter: @djairjr
- Evidence: Debug report + 3 logs provided
Source URL: https://github.com/NousResearch/hermes-agent/issues/61010