AIDiveForge field notes · Evidence recorded October 15–16, 2026
Several prepared tools continued working while external hosts were blocked. That does not make a clean install offline-ready or prove a tool never contacts the network.
A model can be local while dependencies are not
We ran local models in an isolated Docker network and logged external lookup attempts. Some tools tried to contact model catalogs, update hosts, package registries or other services even though the inference model was already local. In several cases, blocking those requests did not prevent the prepared tool from completing its work.
An attempted request is not proof that private input was transmitted, and a successful blocked-network run is not proof that a tool never uses the network. These are bounded observations from particular versions and prepared images. We report the hostnames so another operator can investigate the same startup path, not to infer motives from a DNS lookup.
What the recorded runs showed
| Tool | Observed external hosts or dependency | Result to preserve |
|---|---|---|
| LM Studio / llmster | lmstudio.ai; versions-prod.lmstudio.ai; extensions.lmstudio.ai | Prepared model still loaded and served while blocked |
| LocalAI | index.localai.io; raw.githubusercontent.com; quay.io | Prepared backend served; gallery requests warned |
| Goose | models.dev | Prepared agent kept working while blocked |
| OpenCode | models.opencode.ai; registry.npmjs.org | Hosted provider package was pre-fetched |
| Hermes | First-use Anthropic support from PyPI | Initial offline attempt failed; prepared rerun passed |
| llama.cpp | No DNS lookups observed during its recorded test | Observation covers that run only |
A deployment checklist
For LocalAI, the separately installed backend was part of this inventory. For Hermes and OpenCode, provider-specific support mattered to the hosted reruns. A list of model files alone would have missed those dependencies. An offline requirement should describe which operations must work: installation, startup, inference, tool execution or all of them.
- Inventory the exact tool version, image digest, model files, tokenizer files and provider packages.
- Prepare those artifacts while access is available, recording their origins.
- Start the prepared environment with external access blocked and keep the network and application logs.
- Run the actual tasks, then restart and repeat rather than testing only a warm process.
- Try the documented recovery path in a fresh disposable environment before assuming it works offline.
Keep credentials and hosted comparisons distinct
The hosted agent reruns explicitly permitted the Anthropic API and used the recorded Claude Opus 5.5 model. They demonstrate task outcomes with that provider available; they are not evidence of offline inference. If you compare local and hosted configurations, keep the network policy beside the result.
Our published snapshot records which hosts were denied or allowed. Before applying that information elsewhere, inspect the current tool and your own environment. Package locations, optional extensions and update behavior can change between releases.
What counts as a useful result
A good offline acceptance record contains the prepared-artifact inventory, network boundary, logs, task inputs and the resulting files. If the test fails, keep the failure and identify the missing dependency before retrying. The successful rerun should state what preparation changed.
This approach makes offline operation a repeatable property of your deployment instead of an assumption inferred from the word local. It also gives you a practical way to review updates without guessing whether a first-use download is hidden in the path.
Test conditions and sources
8-vCPU AMD EPYC 9354P cloud VM, each test limited to 3 cores and 10 GB RAM, no GPU, no internet (isolated Docker network). Local-model runs blocked external access. Hosted-model reruns explicitly allowed the Anthropic API. The machine was shared; generation ratios apply to their paired sessions and do not predict GPU or other-machine performance.
These are dated observations, not a claim about the latest release or every machine. Recommendations and example procedures are our interpretation of the recorded results.
- LM Studio: blocked-host observations
- LocalAI: backend and gallery preparation
- OpenCode: provider-package preparation
- Hermes: first-use dependency failure