Going local with localize¶

hlock localize turns any lock (pinned, resolved, or optimized) into its best equivalent that runs entirely on your own machine: a local-capable harness, local models served by a runtime like Ollama or LM Studio, and the same MCP servers, skills and instructions.

$ hlock localize
✓ wrote harness.local.lock  (from harness.lock)
╭─ 🏠 harness.local.lock · localized ─────────────────────────────────────────────────╮
│  harness        opencode  the most preferred local-capable harness installed        │
│  runtime        ollama 0.35.1  http://localhost:11434/v1                            │
│  llm · default  ollama:gemma3:12b                                                   │
│  memory         20.0 GB for weights  (128.0 GB, unified memory (75%))               │
│  candidates     gemma3:12b ✓, devstral:24b, gpt-oss:20b, gemma3:4b ✓, …  (✓ on disk) │
│  evidence       not evaluated (no task packs); ranked by fit, tier, and what's on disk │
╰─────────────────────────────────────────────────────────────────────────────────────╯
  ~ harness claude-code → opencode
  ~ llm default: anthropic:opus → ollama:gemma3:12b
next → hlock sync harness.local.lock --pull  ·  hlock launch --lock harness.local.lock

The source lock isn't touched. -o names the output; harness.local.lock is the default.

What changes¶

The harness, if it's proprietary or can't use a local model endpoint. Each harness profile says whether it's open-source and whether it supports local-models. hlock keeps the lock's harness when it qualifies; otherwise it picks the first installed one of OpenCode, Pi, Codex, Hermes, and Goose. --harness chooses one yourself. A proprietary harness that can use local models, like Claude Code, is allowed with a note. One that can't, like Cursor, is refused.

The models. LLM roles and an LLM decision model on API providers move to a local runtime: --runtime ollama|lmstudio|llama-cpp|…, or the first installed one. Candidates come from the runtime's model catalog and from weights already on disk, and only models that fit this machine's memory are considered: 75% of unified memory on Apple silicon, GPU memory with an NVIDIA card, otherwise half of RAM, or --memory-gb. Ranking prefers stronger agentic models, then weights you already have, then larger ones. The small and decision roles get the smallest capable model. --models a,b replaces the catalog with your own list.

Self-hosted runtimes on your network count as non-local and are replaced, unless you pass --allow-network.

What carries over: MCP servers (written in the new harness's config shape), skills, and instruction files. Remote MCP servers still need the network, so localize flags them; --strict drops them. The old harness's plugins and settings don't carry over, and localize lists them.

The new lock also configures the harness to use the local endpoint. For OpenCode, it adds an ollama OpenAI-compatible provider and sets model = "ollama/gemma3:12b". These settings are merged into the harness's config when you sync or launch.

Evaluate instead of guess¶

With task packs, localize measures the top candidates instead of ranking them. Pass --tasks ./tasks, or run it next to a harness.toml that has [tasks]:

hlock localize --tasks ./tasks

It runs the candidates through the resolver with an all-local search space and records the evidence, next to the source lock's, in the new lock. A code harness runs for real. Agent CLIs run on the backend's adapter for now, and localize says the evidence is a stand-in.

Then¶

The localized lock is an ordinary lock:

hlock sync harness.local.lock --pull      # install what's missing and pull the weights
hlock launch --lock harness.local.lock    # start the local harness

launch keeps each lock's environments separate (.hlock/env/harness.local/ here), so the localized setup never mixes with the original.

Extend the catalogs¶

Model catalogs live in the runtime profiles ([[catalog]]), and the local-endpoint settings live in the harness profiles ([[local-provider]]). Both are plain data you can override; see Harness profiles.