The first update to the MCP server since launch is out now, and the theme this time is servers. Your AI assistant can now change a server’s host-level settings, set up shared file systems for your containers, and show you what’s running on each machine and how much room is left.
What’s new in Cycle MCP 2026.10.14.02
Area | What changed |
|---|---|
Server configuration | New |
Shared file systems | Servers expose shared directories or NFS/EFS mounts, and containers mount them read-only or writable |
Capacity view |
|
Server health | Each server reports agent status, capacity, features and shared file systems |
Telemetry | Instance telemetry compares CPU use to both the host and the container’s own allocation |
Shared file systems
Shared file systems are one of the most useful features of the Cycle platform, and now you can manage them entirely through the MCP.
Containers are usually isolated from each other, and that’s the right default. But some workloads need to share data, and copying large files into every container (or baking them into images) gets slow and expensive fast. Shared file systems solve this at the host level.
There are two kinds:
- Mounts of external storage. A server can mount almost any external file system on the host, such as an NFS share or AWS EFS, and make it available to containers.
- Shared host directories. A server can also expose an ordinary directory on its own disk. Any container on that host can use it if the container opts in.
Each shared file system has an identifier. Containers choose to mount it at whatever path they want, either read-only or writable. Containers that don’t opt in never see it. Every container mounting the same identifier on the same server sees the same files.
A practical example: sharing AI models
Say you’re running training or inference workloads that need a 70GB model. Without shared storage, each container pulls its own copy, and you pay for that in time, bandwidth and disk.
With a shared file system, you can set things up like this instead:
- A downloader service mounts a shared directory as writable. It fetches the model weights (or datasets, or checkpoints) once per host.
- Your training or inference containers mount the same directory as read-only. They read the files directly from the host with no duplication, and they can’t modify or corrupt the source.
The same pattern works well for media libraries, build caches, large reference datasets, or anything else that’s too big to copy around. Because it runs on your own servers, it fits naturally on bare metal hardware with large local drives.
How the MCP handles it
The MCP now manages both halves:
- On the server side,
configure_serveradds, replaces or removes the shared file systems a server exposes. - On the container side,
deploy_applicationandreconfigure_containerhave a newshared_file_systemsfield for mounting an identifier at a path, read-only or writable.
The preview step checks that the two halves line up before anything changes:
- If no server exposes the identifier you’re asking for, the deploy is refused, and the assistant offers to set it up with
configure_server. - If some of the container’s instances would run on servers that don’t expose it, the tool warns you.
So you can just ask “Set up a shared models directory on the GPU servers, give the downloader writable access, and mount it read-only into the training containers” and the assistant handles both sides.
See what’s running and how much room is left
list_servers has a new workloads:true flag that answers a common question: what’s running on this server, and can it take more?
With the flag on, each server lists:
- Its container instances and virtual machines, each with its reserve and limit
- Cycle’s own allocation totals, meaning CPU shares and RAM reserved compared with what the server can allocate
The numbers come from Cycle’s own accounting. They’re the same numbers the scheduler uses, so you don’t have to add them up yourself.
Even without the flag, every server now reports more detail:
- Capacity: CPU threads, cores, physical CPUs, CPU model and total RAM
- Features and shared file systems
- Agent status: online, healthy, shutting down, and last check-in
The assistant also points out servers that need attention, so “Are any of my servers having problems?” now gets a direct answer.
More accurate resource numbers
Several changes make the capacity numbers more trustworthy:
- VMs are counted correctly. Allocation totals now include virtual machines. A VM counts by the reserve of the container that runs it, not by its guest RAM and cores. Guest RAM only counts in full when it’s preallocated. This avoids counting the same capacity twice.
- Better instance telemetry.
summarize_telemetrywithtarget=instancereports CPU use against the whole host and against the instance’s own CPU allocation. It also tells you which server the instance runs on and what its configured reserve and limit are. That makes it much easier to tell whether a container is close to its limit or the whole server is busy.
We also expanded the instructions the MCP gives your assistant. They now explain resource units (10 shares = 1 logical CPU), the difference between reserve and limit, and how shared file systems work. The upshot is fewer wrong answers about capacity.
Getting the update
Cycle MCP is a hosted remote server, so you don’t need to install anything. The new tools and improvements are live for every connected client. If you signed in with write access, configure_server is available right away. If your connection is read-only, the new list_servers details and telemetry improvements still apply. Reconnect with write scope whenever you want to start making changes.
If you’re new to the Cycle MCP, it works with Claude and any other MCP-compatible client. Questions or feedback? Reach out to the Cycle team. We’re always happy to help.
Already a Cycle user and want to connect to our hosted MCP? Check out or docs for details.
Frequently asked questions
What’s new in Cycle MCP 2026.10.14.02?
This release adds the configure_server tool for host-level server settings and support for shared file systems. It also adds a workload and capacity view to list_servers, health status for every server, and more accurate resource and telemetry reporting.
Can the Cycle MCP change server settings?
Yes. With write access, configure_server can change a server’s nickname, tags, placement flags, base volume size, SSH/SFTP access, gateway services and shared file systems. Every change is previewed before it’s applied.
What are shared file systems on Cycle?
They let containers on the same server share files. A server can mount external storage such as NFS or EFS, or expose a directory on its own disk. Containers then opt in and mount it at their own path, read-only or writable.
How can I share large files, like AI models, between containers?
Expose a shared directory on the server and have a downloader container mount it as writable. Then mount the same directory as read-only in the containers that need the files. They all read one copy from the host instead of each downloading their own.
How can I check how much capacity is left on a server?
Ask your assistant to list servers with workloads. The workloads:true flag on list_servers returns every instance on each server with Cycle’s own allocation totals, so you can see reserved CPU and RAM against what the server can allocate.
Do I need to update anything to get the new features?
No. Cycle MCP is hosted, so the update is already live. To use the tools that make changes, your connection needs write scope.