sg-compute.sgit.ai / specs / local_claude

local_claude

A node that runs Claude Code locally against a workspace, provisioned by the Claude_Code__Firstboot user-data section.

Stability
experimental
Capabilities
llm-inference
Typical boot
180s — boot_seconds_typical, from the manifest
Size
1,498 lines — 2.5% of spec code
Family
llm — alongside ollama
create_node
no
Measured
2026-08-24, against repo v0.2.71

create_node does not cover local_claude. EC2__Platform._service_for raises NotImplementedError for everything but docker, podman and vnc. This spec creates nodes through its own CLI and service path, not the control-plane API. The generated CLI verbs below are real; the uniform API is not, for this spec.

The CLI, which nobody wrote

Per-spec CLIs are generated, not written. Spec__CLI__Builder registers the same verb set for every registered spec, so local_claude gets its command surface from the same code that gives every other spec theirs:

# the uniform verb set, for every registered spec
list · info · create · wait · health · connect · exec · delete · ami list|bake · cert

# so, for this one
sg local_claude create --max-hours 1
sg local_claude wait
sg local_claude health
sg local_claude delete

This is the strongest single argument that the thing is a platform rather than a service: adding a spec costs a manifest, a route class and a service — the CLI comes free. The three mechanisms →

What a node of this spec does

Every spec rides the same node lifecycle: a per-node API key minted and written to SSM before launch, composed user-data sections, EC2 tags as the registry, a two-phase health poll, then teardown. Teardown means terminate, not stop — systemd-run --on-active paired with InstanceInitiatedShutdownBehavior=terminate, on a default one-hour timer.

The full node lifecycle, step by step →