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/claude-dev-suite/claude-dev-suite/edge-inferencenpx skills add claude-dev-suite/claude-dev-suite --skill edge-inferencegit clone --depth 1 https://github.com/claude-dev-suite/claude-dev-suiteWrote 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/claude-dev-suite/claude-dev-suite/edge-inference)<a href="https://agentmods.dev/skills/claude-dev-suite/claude-dev-suite/edge-inference"><img src="https://agentmods.dev/badge/skills/claude-dev-suite/claude-dev-suite/edge-inference.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 5.1 | $0.00163 | $0.00662 |
| Opus 5 | $0.00081 | $0.00331 |
| Sonnet 5 | $0.00033 | $0.00132 |
| Haiku 4.5 | $0.00016 | $0.00066 |
Grade A, and why
edge-inference 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 6d 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.
What it actually says
Edge / On-Device Inference
Why on-device (the decision drivers)
- Latency: on-device token/inference in ~tens of ms vs 200–500ms cloud round-trip — decisive for voice, AR, control loops.
- Privacy / offline / cost: data never leaves the device; works without connectivity; no per-call cloud cost.
- Cost of doing it: tight memory/energy/thermal budgets; smaller models; more engineering.
The hardware tiers (match model to silicon)
| Tier | Silicon | Typical model |
|---|---|---|
| MCU / TinyML | Cortex-M + tiny NPU (sub-$1 class) | KB-MB models: keyword spotting, anomaly detection (TFLite Micro) |
| Mobile / AI-PC | Phone NPU, laptop NPU (tens of TOPS) | Quantized 3–8B LLMs, vision |
| Edge box | Jetson Orin/Thor, Coral, Hailo (40+ TOPS) | 7–13B LLMs, multi-camera vision |
Architectural levers
- Quantization is the key enabler: FP16 → INT8 → INT4 trades accuracy for memory/throughput/energy. Most edge LLMs run INT4/INT8. Validate accuracy loss.
- Model choice: small instruct/distilled models (e.g. 3–8B class) and vision models sized to the NPU's memory bandwidth, not just TOPS.
- Runtime: TFLite/LiteRT, ONNX Runtime, ExecuTorch, llama.cpp/Ollama, vendor SDKs (TensorRT for Jetson). Pick what targets the accelerator.
- Budgets: state the TOPS, RAM, and energy-per-inference budget up front — they bound everything. Batch=1, KV-cache memory dominates LLM RAM.
When to recommend edge vs cloud vs hybrid
- Hard latency / offline / privacy / per-unit cost → on-device.
- Large model / variable load / centralized updates → cloud serving.
- Both needs → hybrid (small local model + cloud escalation): see
hybrid-edge-cloud.
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.
- 6d ago First seen · 48 lines · 163 tokens per session scan A 17f0a4fb7397
edge-inference is a skill published in the GitHub repository claude-dev-suite/claude-dev-suite (30 stars, last pushed yesterday), licensed MIT. It adds 163 tokens to every session and 662 once invoked, about $0.0008 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 2026-08-30.
Other skills, from other repositories
integrated-browser
Use this when working on the VS Code integrated browser ("browserView") to understand its architecture and mental model. Covers the embedded Chromium browser, its editor tab, navigation, overlay/layout, sessions, and agent browser tools under src/vs/platform/browserView and src/vs/workbench/contrib/browserView.
spark-environment-setup
Set up a working ML training/inference environment on NVIDIA DGX Spark (GB10, aarch64, CUDA 13). Use when installing PyTorch/Unsloth/TRL/vLLM on DGX Spark, hitting libcudart or wheel-ABI errors on aarch64, or choosing between NGC containers and bare pip installs.
spark-memory-thermal-ops
Manage unified memory and thermals during long-running ML jobs on NVIDIA DGX Spark. Use when planning memory headroom for a training run on GB10, when a job OOMs on unified memory, or when monitoring temperature and power during multi-hour training.
spark-training-gotchas
Preflight and diagnose the ten known failure modes for ML training on NVIDIA DGX Spark. Use when a training run on DGX Spark fails to start, OOMs below the 128GB limit, slows down mid-run, or before any multi-hour training job on GB10.
llama-cpp
Runs LLM inference on CPU, Apple Silicon, and consumer GPUs without NVIDIA hardware. Use for edge deployment, M1/M2/M3 Macs, AMD/Intel GPUs, or when CUDA is unavailable. Supports GGUF quantization (1.5-8 bit) for reduced memory and 4-10× speedup vs PyTorch on CPU.
amc-run-rtsp-calibration
Calibrate a new dataset from live RTSP camera streams via the AutoMagicCalib REST API. Use when the user provides RTSP URLs or asks to calibrate live cameras; VIOS records clips, AMC ingests them, then runs calibration.