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 instructions/dmmdea/offload-harness/claude-mdgit clone --depth 1 https://github.com/dmmdea/offload-harnessWhat 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 | $0.03056 | $0.03056 |
| Opus 5 | $0.01528 | $0.01528 |
| Sonnet 5 | $0.00611 | $0.00611 |
| Haiku 4.5 | $0.00306 | $0.00306 |
Grade A, and why
offload-harness CLAUDE.md scanned grade A with 1 finding 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 2d 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.
Makes network callslowCapability
Not a fault in itself. Listed so you know the mod talks to something, and to what.
| `local-agent` | Coding-agent loop; one-shot, `--queue`, or `--serve` | `127.0.0.1:18800` (serve, OPTIONAL, OFF by default) | shares harness config; flags | `curl 127.0.0.1:18800/v1/models` | How it starts
The opening of the file, as written. The whole thing — 154 lines — stays where its author put it; the contents beside it link to each section on GitHub.
CLAUDE.md — agent orientation map for offload-harness
Local-first Go harness that offloads grunt work (summarize/classify/extract/triage + vision/OCR/STT/
media-gen) to a free Gemma-4 cascade on llama.cpp. Ships as a CLI, an MCP server
(local-offload), and an optional local coding agent (local-agent). The cascade never
calls cloud; on low confidence it returns a structured defer and the caller does the task.
(offload_nim is the one remote surface — an explicit, caller-invoked side channel that nothing
escalates or falls back into.) Every command below was executed successfully while writing this file.
Components & ports
| Binary / service | Role | Port | Config | Health check |
|---|---|---|---|---|
llama-swap (fronts llama.cpp) |
Serves the model tiers | 127.0.0.1:11436 |
$OFFLOAD_HOME\llama-swap.yaml |
local-offload doctor → health: OK |
local-offload |
Offload CLI + MCP (stdio) | — (talks to :11436) | ~/.local-offload/config.json |
local-offload doctor |
local-agent |
Coding-agent loop; one-shot, --queue, or --serve |
127.0.0.1:18800 (serve, OPTIONAL, OFF by default) |
shares harness config; flags | curl 127.0.0.1:18800/v1/models |
| OpenWebUI | Chat GUI over the agent (OPTIONAL) | 127.0.0.1:8081 |
env in scripts/openwebui-stack.sh |
curl 127.0.0.1:8081/health |
The --serve endpoint is unauthenticated and drives write/GitHub tools → loopback-only;
--listen-trusted-network is required to bind beyond loopback (loud warning).
Model tiers (served by llama-swap on :11436)
| Alias | Config key | Role |
|---|---|---|
gemma4-e2b |
triage_model |
Fast entry tier — triage / classify. |
offload-e4b (alias gemma4-e4b) |
model |
Workhorse — summarize / extract; default agent planner. |
gemma4-26b-a4b |
escalation_model / reasoning_model |
MoE tier tried before deferring. |
embeddinggemma |
(memory stack) | Embeddings. |
qwen3vl-4b, whisper-stt, whisper-stt-hq |
vision_model / stt_model[_hq] |
Vision + speech. All opt-in and DERIVED: a tier declares a media_seats entry, which renders the llama-swap seat and writes the binding together. No seat, no binding, route defers. |
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.
- 2d ago First seen · 154 lines · 3,056 tokens per session scan A fa264cb3f128
offload-harness CLAUDE.md is an instructions file published in the GitHub repository dmmdea/offload-harness (2 stars, last pushed 3d ago), licensed Apache-2.0. It adds 3,056 tokens to every session, about $0.0153 per session on Opus 5. A static security scan graded it A with 1 finding (makes network calls). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other instructions, from other repositories
maclocal-api CLAUDE.md
Instructions for scouzi1966/maclocal-api, covering claude code reference for maclocal api, project overview, project structure, provider source ownership and dwarfstar dependency boundary.
docmind-ai-llm AGENTS.md
Instructions for BjornMelin/docmind-ai-llm, covering docmind ai: agent instructions, purpose, layout, quick commands with uv and non-negotiables for ci and security.
Bonsai-demo AGENTS.md
Instructions for PrismML-Eng/Bonsai-demo, covering agent guide — tuning the bonsai demo, why the 27b models (what to show off), the models, knobs that matter (27b) and adding mcp servers.
James-RAG-Evol CLAUDE.md
Instructions for Hashevolution/James-RAG-Evol, covering james — session briefing for claude code, what james is, where we are right now, critical rules for this session and where to look next.
Local-Model Harness Tuning Workflow
Pragmatic playbook for running the harness against local LLMs (2B-13B class) and iterating until numbers improve. Captures gotchas that bite every time.
myrmion-graphrag CLAUDE.md
Claude Code instructions for e2its/myrmion-graphrag, covering claude.md, qué es esto, herramientas del servidor myrmion-graphrag (documentos), herramientas del servidor myrmion-codebase (código) and mantenimiento obligatorio del codebaseinventory.