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/kerbelp/metatron/claude-mdgit clone --depth 1 https://github.com/kerbelp/metatronWrote 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/instructions/kerbelp/metatron/claude-md)<a href="https://agentmods.dev/instructions/kerbelp/metatron/claude-md"><img src="https://agentmods.dev/badge/instructions/kerbelp/metatron/claude-md.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 | $0.01156 | $0.01156 |
| Opus 5 | $0.00578 | $0.00578 |
| Sonnet 5 | $0.00231 | $0.00231 |
| Haiku 4.5 | $0.00116 | $0.00116 |
Grade A, and why
metatron CLAUDE.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 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.
How it starts
The opening of the file, as written. The whole thing — 99 lines — stays where its author put it; the contents beside it link to each section on GitHub.
CLAUDE.md
Ground rules for working in this repo. Read this first.
What Metatron is
Metatron is a self-hosted system that captures a company's real implementation decisions — preferred patterns, rejected approaches, edge cases, internal conventions — as structured decisions. In the primary, files-first mode they live as git-tracked OKF markdown consulted directly by coding agents; an optional MCP (Model Context Protocol) server layer serves the same decisions over the wire. The goal is for an agent to write code like a senior engineer who already knows this codebase. It runs against a private codebase, so assume sensitive data and on-prem deployment.
Tech stack (decided — not open for debate)
These are locked. Do not re-litigate or substitute them:
- Python 3.12+
- Official MCP Python SDK for the server
- tree-sitter for language-agnostic code parsing
- SQLite for the decision store — but always behind a storage interface, because the schema must be portable to Postgres later
- pytest for tests
- uv for dependency management
Core principle
- Decisions are stored as structured records (fields: pattern, context/scope, rationale, confidence, source refs) — never as prose specs.
- Nothing enters the canonical set without human curation. No decision self-promotes. Crossing the canonical boundary — promote, demote, reject — is always human-gated. This invariant is absolute.
Source of truth (depends on deployment)
Metatron supports two ways of operating, and the source of truth differs:
- Files-first mode (primary/default). The git-tracked OKF files are the source
of truth: the directory (
candidate/vsdecisions/) is the status, decisions are curated as plain files reviewed via pull request (default review gate:pr), and the database is a derived, rebuildable serving index (metatron mirror import). The curation UI mounts the bundle directly (metatron ui --files). - MCP / database mode (optional serving layer). SQLite is the source of truth;
the git-tracked OKF markdown bundle is a synced mirror (
metatron mirror sync). Curation happens in the store (CLI/UI), and decisions are served to agents over MCP with relevance ranking and the feedback loop.
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 · 99 lines · 1,156 tokens per session scan A f38f2a9443b6
metatron CLAUDE.md is an instructions file published in the GitHub repository kerbelp/metatron (24 stars, last pushed 1mo ago), licensed MIT. It adds 1,156 tokens to every session, about $0.0058 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 instructions, from other repositories
thinkrail AGENTS.md
AGENTS.md instructions for JetBrains/thinkrail, covering thinkrail, module structure & boundaries (top-priority requirement), engine: pi only, in-process, architecture (three rings) and repo layout.
vant AGENTS.md
AGENTS.md instructions for dhaupin/vant, covering agents.md — vant agent guide, what is vant?, your job, brain router interface and middleware chain.
lihongwei-cn CLAUDE.md
Claude Code instructions for LiHongwei-cn/lihongwei-cn, covering claude.md — claude code 项目记忆, 用户身份, 回复风格（红线）, 网站 and 技术栈.
cortex copilot-instructions.md
Instructions for CortexPrism/cortex, covering workspace: cortexprism, project: cortex, workspace id: 82f5beed-6cea-481f-924c-f8c086994250, contextstream rules and required tool calls.
thinkrail CLAUDE.md
Claude Code instructions for JetBrains/thinkrail, a project described as: Vibe code with pi in a lightweight, real IDE - The Vibe You Need.
Warden-AI copilot-instructions.md
Copilot instructions for rynald0cst0ltziam/Warden-AI, covering warden — context governance and verification layer, session start (important — do this first), layer 1: before starting work — context selection, layer 2: during work — tool output pruning and enforcement hooks (automatic).