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/nexaduo/marc/agents-mdgit clone --depth 1 https://github.com/NexaDuo/mARCWhat 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 | $0.02461 | $0.02461 |
| Opus 5 | $0.01230 | $0.01230 |
| Sonnet 5 | $0.00492 | $0.00492 |
| Haiku 4.5 | $0.00246 | $0.00246 |
Grade A, and why
mARC AGENTS.md 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 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.
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 — 143 lines — stays where its author put it; the contents beside it link to each section on GitHub.
AGENTS.md
mARC (Multi-Agent Relay Control) is a Claude Code plugin + self-marketplace
that packages a portable, cross-repo AI engineering team: @techlead (channel
operator) convenes specialists @dev, @sre, @design, @sec, @research. The
team's
governance travels via one plugin; each consuming repo keeps its own facts.
This repo is special: it is BOTH the product's source AND where the team dogfoods on itself. Self-improvement lessons legitimately flush to source here — the one context where that is allowed (see Constraints).
Architecture
- Plugin lives at
harnesses/claude-code/marc/; the marketplace manifest is at repo root.claude-plugin/marketplace.json(namenexaduo, pluginmarc→ installmarc@nexaduo, invoke/marc:tech-lead). - Leaders = skills (
skills/<leader>/, invoked/marc:<leader>); specialists = a shared flat pool (agents/*.md). Any leader convenes any specialist. - Nesting under
harnesses/<harness>/reserves the namespace for future non–Claude-Code harnesses. The growth model is indocs/ARCHITECTURE.md— don't duplicate it here. core/is the editable source;harnesses/*/marc/(skills, agents) is compiled output — never hand-edit it. Runscripts/compile_prompts.pyto regenerate fromcore/+ each harness'scompile.json, and commit the regenerated files.
Operating principles
These distill a comparative study of agent-orchestration frameworks (OpenHands, MetaGPT, CrewAI, Claude Code Agent Teams) down to what holds for mARC:
- Stateless dispatch / context hygiene. Specialists run fresh per task; durable state lives in git, the board/issues, and this file — not in long conversations.
- Spec-driven leverage. A vague task propagates destructively across parallel agents. The tech-lead's main leverage is spec clarity + acceptance criteria before dispatch (the sufficiency gate).
- Dispatch in the background; isolate parallel writers. Synchronous dispatch blocks the operator channel — fan out in background and track via notifications. When specialists edit files concurrently, isolate them (git worktrees) to avoid clobbering.
- Bounded loops. Cap runaway debugging; stop and surface rather than burn tokens in an infinite fix loop.
- Tiered, size-capped memory with absolute decay. Local/session memory is
bounded (≤ 200 lines / ~2 KB) and organized as a lightweight recall index
separating permanent invariants (
[PINNED]) from time-bound entries with absolute dates ([EXPIRES: YYYY-MM-DD]). Substantive findings or oversized bodies spill into PR-gated artifacts (docs/marc/or consumer workspace) and are fetched on demand, never loaded unconditionally. - No self-merge; independent review. Every PR gets a security pass; the author can't self-approve.
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 · 143 lines · 2,461 tokens per session scan A bd669de17c8c
mARC AGENTS.md is an instructions file published in the GitHub repository NexaDuo/mARC (6 stars, last pushed 3d ago), licensed MIT. It adds 2,461 tokens to every session, about $0.0123 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 instructions, from other repositories
kungfu copilot-instructions.md
Copilot instructions for kungfu-systems/kungfu: Read and follow AGENTS.md before proposing or running changes. It is the shared agent router for this repository, including the canonical Shifu development and build entrypoint.
sofagent GEMINI.md
Instructions for KongFangXun/sofagent, covering gemini.md · sofagent gemini cli 适配（薄挂载）, 会话开始时按序加载（四层加载链）, 审计强制（平台无关） and 连接 mcp server.
agents CLAUDE.md
Instructions for eloylp/agents, covering claude.md, project overview, directory structure, config model and build & run.
ideate AGENTS.md
Instructions for paultyng/ideate, covering agents.md, build features as mcp tools first, ui second, when to add ui and pattern.
growmos AGENTS.md
Instructions for codician-team/growmos: This repository keeps a knowledge graph in .growmos/ (entities, typed relations, provenance, profiles, a journal). It is the shared world model that survives context windows. Treat it as memory you read at the start of work and write to as you develop. Zero-config commands.
omk CLAUDE.md
Claude Code instructions for dmae97/omk, covering claude.md, precedence (read this first), omk runtime map (what you're operating inside), this repo: open-multi-agent-kit / omk monorepo and build / test / check commands.