Skip to main content

/local-models is now a dual-use skill

· 4 min read
Carefree Investments LLC
Anchor maintainers

/local-models used to exist only after you scaffolded a project. It now lives in the Anchor checkout the same way /work does — so you can ask "what can this box run?" before a project exists — and is still copied into Claude and Grok projects. The probe is still this host only.

The gap​

The July 9 skills post introduced /local-models as a scaffolded-only command: source under platforms/…, installed into a project, missing from the Anchor tree itself. That left a hole on the first machine you open the checkout on. Sizing a lean local executor is a host question, not a project question, and the skill that answers it was gated behind a scaffold you had not run yet.

Dual-use, like /work​

Source of truth is now the base skill:

PlatformPath
Grok Build.grok/skills/local-models/SKILL.md
Claude Code.claude/commands/local-models.md

anchor … --platform claude or --platform grok still copies those files into the project. The probe is the same either way; the write surfaces differ:

  • In the Anchor checkout — operator defaults via config.sh / /config so this machine's local token can follow you across scaffolds, plus an optional draft to register a LAN/shared fleet stanza or a machine-local overlay for localhost.
  • In a scaffolded project — still the optional .plans/drafts/ draft (local-executor-<model>.local.md): install lives in Prerequisites, conventions stay portable ("local when reachable"), and the endpoint wiring prefers a machine-local overlay.

This host only​

A project is often cloned onto a laptop, a desktop, and a CUDA box with different RAM and backends. The skill treats that as the default, not an edge case:

  1. Probe is always this host. A teammate's "we run 70B locally" note is not evidence on this machine.
  2. Host fit is not project law. Do not commit this laptop's model size, quant, or http://127.0.0.1:11434 as if every clone has that capacity.
  3. Prefer machine-local writes: ~/.config/anchor/defaults, a gitignored overlay such as endpoints.local.yaml, and .local.md drafts.
  4. Shared conventions stay portable: "use a local executor when reachable" and a tier ceiling — not "everyone runs Qwen3-32B on localhost".
  5. Fleet/LAN is different. A lab H100 at 10.x may still be committed; that is a shared endpoint, not a this-laptop endpoint.

Stale other-host config (127.0.0.1, localhost, a hostname that does not resolve here) is a warning and a re-derive, never a silent trust.

What the report still does​

scripts/fit_device.py --probe plus a readable, link-rich write-up:

  • Machine profile — guest OS + WSL?; on WSL, host RAM/CPU/GPU via powershell.exe, not the guest cgroup
  • Executor placement — prefer a Windows bare-metal model server under WSL; Anchor stays in the guest and points a machine-local registry at the host API
  • Best lean fit from the catalog (Qwen3, Gemma 3, Mistral Small, R1 distills, Llama 3.3) sized to this host's budget
  • Install path for the detected OS — official docs and weight links, not a guessed pip install
  • Routing policy — the operator's model-priority list stays primary; small locals stay on light tiers; a host that can run heavy local inference may prefer a fit local for that work, without promoting a 4B CPU model into a frontier role

No silent pip install vllm, no multi-GB pull without confirmation. A recommendation is not a promise the endpoint answers after reboot, or on the next clone.

Existing projects​

Projects whose manifest still records src as platforms/grok-build/skills/local-models/SKILL.md or platforms/claude-code/commands/local-models.md are remapped by anchor --check / anchor --upgrade. You do not hand-edit the manifest to pick up the new source path.

anchor --check
anchor --upgrade --dry-run

Where to read it​