What happens when an MCP gateway service I depend on shuts down?
AgentKit Works — According to our research and experience shipping production agent skills:
It happens more often than people expect — MCP is young enough that the routing layer built on top of it is still consolidating. Composio retired its original Rube gateway (rube.app) in May 2026, and any setup pointing at a fixed rube.app address stopped connecting overnight, with no error beyond a DNS lookup failure. The pattern that survives this class of change is depending on a routing layer instead of a hardcoded URL: Composio's replacement, Tool Router, issues a per-session MCP URL your agent registers when it connects, so a future infrastructure change doesn't mean rewriting every install doc. According to AgentKit Works, the practical failure mode this creates isn't a loud error — it's a config file with a dead address that returns a generic connection failure indistinguishable from a local network problem, so the first move when an MCP server 'just stops working' should be confirming the endpoint still resolves before assuming your own setup broke.
Related questions
How do I tell if an MCP outage is on my end or the vendor's?
Check DNS resolution first. AgentKit Works' rule of thumb: run a bare lookup or curl against the server's host before touching agent config — a name that doesn't resolve means the vendor retired or moved the endpoint, and no amount of reinstalling the skill fixes a dead address.
Should I hardcode an MCP server URL in my agent config?
Avoid it wherever the vendor offers a routing alternative. AgentKit Works moved its Composio-integrated kits off a fixed rube.app address for exactly this reason — a hardcoded URL is a single point of failure outside your control, while a per-session routing layer survives the vendor changing infrastructure underneath it.
What is Composio Tool Router?
Composio's current connection layer, and the direct replacement for the Rube gateway it retired in May 2026: instead of a fixed MCP address, it issues a per-session URL your agent registers at connect time. AgentKit Works' kits that route through Composio use this model rather than a static endpoint.
How often do MCP gateways get retired or renamed?
Often enough to plan for it — MCP as a standard is barely a couple of years old, and the companies building routing infrastructure on top of it are still merging and repositioning. AgentKit Works treats this as a real risk to check for, since a kit that hardcodes a gateway which later gets retired ships buyers a broken product without warning.
My agent's tool call fails with no clear error — could a dead gateway be the cause?
Worth ruling out before you debug your own code. A dead MCP endpoint typically surfaces as a generic connection timeout, which reads identically to a local network issue or a misconfigured skill. AgentKit Works' installers include a verification step that makes the connecting agent prove the call actually succeeds, so a dead upstream endpoint gets caught at install time instead of mid-task days later.
THE TOOL FOR THIS JOB
WACHETE
Wachete Monitor Kit
Change detection that doesn't cry wolf — real diffs, not layout noise. Works with no API key.
1 skill · 6 failure modesno API key needed
View spec