Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add skills/dmmdea/offload-harness/local-offload-setupnpx skills add dmmdea/offload-harness --skill local-offload-setupgit clone --depth 1 https://github.com/dmmdea/offload-harnessWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/dmmdea/offload-harness/local-offload-setup)<a href="https://agentmods.dev/skills/dmmdea/offload-harness/local-offload-setup"><img src="https://agentmods.dev/badge/skills/dmmdea/offload-harness/local-offload-setup.svg" alt="Measured on agentmods" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00204 | $0.00828 |
| Opus 5 | $0.00102 | $0.00414 |
| Sonnet 5 | $0.00041 | $0.00166 |
| Haiku 4.5 | $0.00020 | $0.00083 |
Grade A, and why
local-offload-setup scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 3d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 59 lines — stays where its author put it; the contents beside it link to each section on GitHub.
local-offload-setup
This skill is a thin wrapper. The canonical, always-current procedure lives in the repo at
setup/SETUP-AGENT.md — follow that, not a copy here, so the skill can never drift from the
scripts it drives.
Hardware gate
| Detected | Backend | Notes |
|---|---|---|
| NVIDIA GPU (~8 GB VRAM) | cuda |
First-class. nvidia-smi or CIM name match selects it. |
| AMD Radeon, incl. RDNA3 iGPU (780M / gfx1103) | vulkan |
Native Windows Vulkan. ROCm/HIP and WSL2-for-iGPU are research-verified dead ends — do NOT attempt. |
| No GPU | cpu |
Fallback. Slow (<8 t/s); 26B tier needs ≥48 GB RAM. |
| RAM < 32 GB | (any) | Install E4B workhorse only: OFFLOAD_WITH_FAMILY=0. |
| Free disk < 25 GB | — | Hard blocker; detect.ps1 exits 1. |
Procedure
Run the three scripts in order, reading each one's final JSON line, and follow
setup/SETUP-AGENT.md for the decision tree at every branch:
pwsh -NoProfile -File setup\detect.ps1 # -> backend + blockers
pwsh -NoProfile -File setup\install.ps1 # -> pinned binaries, models, Go build, rendered yaml
pwsh -NoProfile -File setup\selftest.ps1 # -> receipt JSON: verdict pass|warn|fail
Then start the stack and register the MCP (Step 3 of setup/SETUP-AGENT.md):
& "$env:OFFLOAD_HOME\llama-swap\llama-swap.exe" --config "$env:OFFLOAD_HOME\llama-swap.yaml" --listen 127.0.0.1:11436
claude mcp add local-offload -- "$env:OFFLOAD_HOME\harness\local-offload.exe" mcp
Do not
- Do NOT install CUDA or ROCm — the pinned llama.cpp release binaries carry their own runtime; Vulkan uses the GPU driver.
- Do NOT substitute a different model, build, or "latest" asset if a pin 404s or a hash mismatches — stop and surface it.
- Do NOT update a GPU driver, reboot, or start the unauthenticated
local-agent --serveendpoint beyond loopback without asking the human first.
For operation after install (chat, offload tasks, driving the coding agent, diagnosis), see
CLAUDE.md and docs/OPERATOR-GUIDE.md in the repo.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 3d ago First seen · 59 lines · 204 tokens per session scan A 382b5496ffa0
local-offload-setup is a skill published in the GitHub repository dmmdea/offload-harness (2 stars, last pushed 4d ago), licensed Apache-2.0. It adds 204 tokens to every session and 828 once invoked, about $0.0010 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
implementing-llms-litgpt
Implements and trains LLMs using Lightning AI's LitGPT with 20+ pretrained architectures (Llama, Gemma, Phi, Qwen, Mistral). Use when need clean model implementations, educational understanding of architectures, or production fine-tuning with LoRA/QLoRA. Single-file implementations, no abstraction layers.
swift-mlx-lm
MLX Swift LM - Run LLMs and VLMs on Apple Silicon using MLX. Covers local inference, streaming, wired memory coordination, tool calling, LoRA fine-tuning, embeddings, and model porting.
swift-mlx
MLX Swift - High-performance ML framework for Apple Silicon with lazy evaluation, automatic differentiation, and unified memory.
test-afm-assertions
Run AFM assertion tests across one or more models with deterministic pass/fail validation. Tests tool calling (including XML format), stop sequences, logprobs, think extraction, streaming, prompt cache, error handling, and performance.
add-afm-model
Add support for a new HuggingFace MLX model to AFM. Use when user wants to add, onboard, or check compatibility of a model — handles everything from "already supported" to implementing new architectures.
afm
Maintain and extend AFM (maclocal-api), a Swift OpenAI-compatible local LLM server and CLI for Apple Foundation Models, MLX models, API gateway proxying, and Vision OCR. Use when working on AFM CLI commands (afm, afm mlx, afm vision), OpenAI /v1/chat/completions and /v1/models behavior, streaming SSE, tool-calling…