Fleetix is the generic fleet product: a typed Pkl schema for fleet-wide host, link, GPU, domain, and service topology, Rust bindings and a standalone CLI over it, and NixOS and Home Manager modules that expose a generated topology sidecar to a consumer flake. A consuming flake owns the real data; Fleetix owns the shape, the bindings, the export path, and the reusable runtime pieces that frontends build on.
Every fleet ends up with its own ad-hoc map of hosts, addresses, VPN profiles, GPU routes, and service placements, plus per-consumer copies of the code that reads it. That duplicates validation, makes derived answers (which address do I use for this host, which GPU belongs to this route) inconsistent between Nix and Rust, and encourages a second deployment engine inside whichever consumer gets impatient first.
- Data and schema are separate. Fleetix ships the Pkl schema shape and a synthetic example; production topology stays in the consumer's flake, and generated sidecars are the consumer's artifacts.
- One derivation path for Nix and Rust. The sidecar converts Pkl to Nix, and the Rust crate reads the same topology, so both sides answer from the same data.
- Typed results, typed errors. Libraries return typed results; frontends own parsing and presentation.
- Independent of any one fleet. Fleetix takes explicit topology, source, state, tool, and policy inputs instead of assuming Canix's paths, users, repositories, or fleet names.
- Publish as a product, not as source distribution. Public Rust libraries are published to crates.io by Simit CI and consumed as versioned Cargo dependencies; flake inputs serve Nix modules and packaging. A consumer must never reuse a library by invoking another product's CLI.
- One generic engine. canix-toolbelt builds architecture conventions on Fleetix, and specialist engines (Nix management, secrets, Crossbow, Harbor) keep their ownership rather than being absorbed.
- MCP authorization is explicit.
Server::newdenies every invocation until the embedding application installs an authorizer, tool annotations and argument names are not authorization, and privileged tools must not get an unconditional allow policy. - The optional MCP server is composable: consumers register their own typed tools beside Fleetix's, duplicate names are rejected, and handlers capture their own dependencies rather than invoking a downstream executable.
The generic topology schema (hosts, links, domains, GPUs, services), accessors and ordered address-policy resolution, validation, export and sidecar generation, routing and management candidates, the build-train construction slice, paired dev/prod service surfaces with exact-output promotion, MCP client reconciliation, and the composable Rust MCP server. Agents use the shipped MCP server to read a selected topology snapshot — hosts, links, and derived addresses — over stdio instead of parsing Pkl files.
Fleetix is not Canix, does not carry production fleet data, and is not a mandate to absorb every neighbouring crate. General whole-host deployment execution remains an extraction target: publication, target transfer, live resource admission, and activation stay caller-owned until their deployment slices are extracted. It does not promise zero-downtime or host-HA through service surfaces.
The recent work is the paired service-surface backend (independent dev and prod
service guests, immutable-output promotion, guarded recovery through a separate
fleetix-surfaces frontend) and generic topology routing that resolves ordered
candidates against a caller-supplied policy and probe. Immediately before that,
hosts gained free-form tags and a worker device type for headless compute.
Canix pins Fleetix as a flake input and reads topology through its MCP tools;
canix-toolbelt builds its conventions on Fleetix; the fleet superproject
qualifies Fleetix revisions with the other components; and downstream products
such as harbor-cad consume its GPU route helpers.