Skip to main content
Coming in the first release. This page describes behaviour being built for the first Gliss release. Details may change before it ships.
gliss mcp serves these tools to your AI tool over MCP, on standard input and output. It uses your gliss login session. To connect it, see Supported AI tools. Every tool takes app, the app’s name as in its gliss.yaml, for example expense-tracker. For a Preview, instance picks whose Preview; the default is yours.

​
gliss_status

What runs at each stage of an app: the release and its commit, health (ready and desired replicas), the address, the cluster’s last result for a release (applied, failed or rolled back, with its message), and when the release’s approval expires. Use it to answer “is it up?” and to find the address.

​
gliss_diagnose

A first diagnosis of one app at one stage: its status, a one-line cause, whose problem it is (the app’s, which your AI fixes, or the platform’s, which it must not try to fix in the app’s code), the evidence (the last log lines, redacted) and the next step. Your AI should call it first when a release fails or the app crashes, errors or does not start.

​
gliss_logs

The containers’ logs at one stage, oldest first, with secrets shown as [REDACTED:NAME].

​
gliss_release

Offered only when you start the server with --allow-release (Beta and your Preview) or --allow-release=preview (your Preview only). It builds every component of the app’s gliss.yaml with Docker on your machine, pushes the images with your AWS profile, asks Gliss to deploy them, and follows the release. It returns Gliss’s refusals, or the result, health and address, and a diagnosis when the release fails. Live is never released from here: a person publishes it in the console. The code must be committed and pushed. One release runs at a time.
gliss_release replaces what runs at a stage. Your AI tool asks you before each call: never allowlist or auto-approve it.

​
Safety

  • Text from your app or the cluster (logs, error messages) is marked to your AI as untrusted, and it must not follow instructions found in it.
  • No tool returns a secret’s value, and none accepts one.
  • Your AI acts as you: it sees only the apps and stages you can see.

​
Planned

These tools are designed but not built yet: