litmus CLAUDE.md

Repository guidance for Litmus, an open-source test harness that evaluates MCP servers and records proof of results on the Base blockchain. An MCP server is a program that provides tools to AI agents.

In plain words
What is it for?
Understanding the repository layout, developing the test and grading packages, handling on-chain proof code, and avoiding secrets or private operations configuration.
Why use it?
It tells an agent which files and external specifications are authoritative and identifies information that must never be added to the public repository.

Instructions file

Install

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.

agentmods
npx agentmods add instructions/polygraphso/litmus/claude-md
Clone the repo
git clone --depth 1 https://github.com/polygraphso/litmus
Per session 1,877 This file is loaded in full into every session.
When invoked 1,877 The same file — it is already loaded in full.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

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.

ModelPer sessionOnce invoked
Fable 5 $0.01877 $0.01877
Opus 5 $0.00938 $0.00938
Sonnet 5 $0.00375 $0.00375
Haiku 4.5 $0.00188 $0.00188

Measured 2d ago against content hash 3fd287dbcb7d, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

litmus 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 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.

CLAUDE.md · 106 lines

How it starts

The opening of the file, as written. The whole thing — 106 lines — stays where its author put it; the contents beside it link to each section on GitHub.

litmus — workspace context for Claude

This repo is the open @polygraphso/litmus harness — the behavioral litmus test for MCP servers, its onchain proof (EAS attestation on Base), and the agent-gate. It is the engine; the hosted, operator-run grading service that runs it lives in a separate private repo and is not here.

This repository is public. Every commit, file, and history entry is world-readable. Never add secrets, credentials, internal hostnames, customer data, ops/infra config, the runner, or .env files. .env.example carries config shape only. Openness is the point: the grade is reproducible precisely because the harness is public, so anyone can re-run it and disprove a false grade.

Where to look first

  • README.md — what this repo is, the layout, build/test/release.
  • packages/litmus/README.md — the npm-facing usage docs.
  • The methodology + proof spec live at polygraph.so (the prose docs are not kept in this repo). The open source here — packages/probes (the probes/grading) and packages/onchain (the EAS schema/encode) — is the source of truth; anchor against the code and the published methodology, don't relitigate them.

What we're building (one paragraph)

A harness connects to an MCP server like an agent would — stdio for local packages, Streamable HTTP for remote URLs — fingerprints the exact tool surface (tools/list → canonical JSON → sha256 → bytes32), then runs four probe categories: C-01 tool-output injection, C-02 permission overreach (egress, in a hardened default-deny Docker sandbox with a sinkhole), C-03 sensitive-data handling (planted canaries), C-04 adversarial-input handling (malformed/oversized + jailbreak inputs; fails on crash, internals-leak, or amplification → caps at D). It grades A–F and produces a deterministic, content-addressed evidence bundle. This package grades and verifies only — it does not mint (no pin-to-IPFS, no EAS write, no /mint hand-off). packages/onchain keeps the EAS read/encode/decode + the agent-gate, so an agent can still verify a published attestation; producing one is out of scope. The grade is reproducible — the harness is open and deterministic, so anyone can re-run it against the same server and disprove a false grade.

Read the full file on GitHub · 106 lines

Changes

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.

  1. 2d ago First seen · 106 lines · 1,877 tokens per session scan A 3fd287dbcb7d

Subscribe to this mod's changes

litmus CLAUDE.md is an instructions file published in the GitHub repository polygraphso/litmus (8 stars, last pushed 1mo ago), licensed Apache-2.0. It adds 1,877 tokens to every session, about $0.0094 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.