Skip to content

The Durability Curve · free by email

Every new piece as it ships, and the Weekly on Saturday, with the calls we grade.

107Pieces, dated
1245Claims verified at the primary source
7Instruments, no login

Your address goes to Substack, where you confirm it. Nothing is stored on this site.

The latest issue · Anthropic Says It Has Met Its Own AI Research Threshold. NVIDIA's New Chief Challenged Huang's Bets. Read it first

Use this site from inside your AI assistant

Add one address to Claude, ChatGPT, Claude Code or Cursor and it can check whether your coding agent's own guard settings hold, against the runs measured for this site's kits. It can also search every essay here, read the claim ledger behind each checked essay, look up every call made in the Weekly, and tell you when an AI model came out, what its API costs and whether the free plan has it, each with a dated source. It can only read. There is no account and no sign-in.

01 · Address

No account · no sign-in

Your connector address

https://mcp.durabilitycurve.com/mcp

02 · Setup

Add it to your assistant

Claude, on the web or the desktop app

  1. Open Customize, then Connectors.
  2. Click Add, then Add custom connector.
  3. Name it The Durability Curve, paste your address, set Authentication to No sign-in and click Add.
  4. In a chat, open the + menu, choose Connectors and switch it on.

Free, Pro and Max accounts can do this; a free account can add one custom connector. On a Team or Enterprise plan, an owner adds it under Organization settings, then Connectors. Once added on the web, it also works in the Claude phone apps.

ChatGPT, on the web

  1. Go to chatgpt.com/plugins and click +, then Create custom MCP server.
  2. Name it The Durability Curve, paste your address and choose No authentication.
  3. Read the warning, confirm it, and click Create as a plugin.
  4. In a chat, type @ and pick it.

This needs a Plus or Pro account; free ChatGPT accounts cannot add custom servers yet. ChatGPT warns about any server it has not reviewed itself, this one included. Business and Enterprise workspaces need an admin to allow it.

Claude Code

claude mcp add --transport http durability-curve https://mcp.durabilitycurve.com/mcp

Then type /durability-curve:check_my_agent_setup and Claude Code reads its own settings, removes any secrets and checks them. It changes nothing without asking you.

Codex

codex mcp add durability-curve --url https://mcp.durabilitycurve.com/mcp

Then ask: check my Codex setup with The Durability Curve.

Cursor and VS Code

Add to Cursor Add to VS Code

Each button opens the app and asks you to confirm before anything is added. Or add it by hand. Cursor: add this to ~/.cursor/mcp.json.

{
  "mcpServers": {
    "durability-curve": {
      "url": "https://mcp.durabilitycurve.com/mcp"
    }
  }
}

VS Code with Copilot: add this to ~/.copilot/mcp-config.json. Keep it in this file in your home folder rather than in a project, so your address is never committed with your code.

{
  "mcpServers": {
    "durability-curve": {
      "type": "http",
      "url": "https://mcp.durabilitycurve.com/mcp"
    }
  }
}

03 · Prompts

6 to try

Things to ask it

Read my Claude Code settings and hook scripts, and check them with The Durability Curve: would my deny rules stop a force-push?

Search The Durability Curve for what it says about agent reliability, then show me the claim ledger behind the best match.

Which calls in The Durability Curve Weekly resolve before April 2027, and at what confidence?

Read the essay about the marathon gap and tell me which of its claims were checked against a primary source.

When did Claude Sonnet 5.5 come out, what does its API cost, and is it on the free plan?

Here is a paragraph I am reading. Has The Durability Curve written about this already?

04 · Limits

Read only

What it can and cannot do

It has eight tools, and each one only reads. check_agent_setup takes the settings files your assistant reads from your setup (permission rules, hooks, skills and sandbox settings for Claude Code, Cursor or Codex) and says which guards would not hold, each finding with the measured run behind it and the kit that tests it on your machine. It runs nothing and is not a security audit: a finding is a prediction from your settings' text and our runs, and no finding does not mean a guard holds. Without a code it shows how many problems it found and the first one in full; with the paid code it shows every finding and the settings that held. get_kit lists the kits, and with the paid code gives your assistant one to run.

The other six read this site: search across the essays, their claims, the weekly calls, the five laws the essays are tested against, the free tools and the AI model record; get_essay for the public text of one essay, whole or a section at a time; get_claim_ledger for the claims one essay makes and how each was checked; get_weekly_calls for every call with its printed confidence; get_ai_models for when each model came out, its API price and who can use it, from the release record and the Access Gap; and related_to, which takes a passage you paste and finds what has been published here on the same subject. For a paid essay it sees only the free part, the same as any visitor.

It retrieves what is published here and says where each answer came from. It does not decide whether a claim is true. A claim ledger records what the source said and how it was checked, and your assistant should cite the page it links to.

It sees what your assistant sends it, such as a search phrase or a settings file, uses that to answer, and keeps none of it. It tells your assistant to remove tokens, keys and env values from a settings file before sending it, and drops any it is sent anyway. What it keeps is a weekly record for each connector code: the week, which tools were used and how often, under a scrambled form of the code that cannot be turned back into it. The privacy page explains it, and how to have yours deleted.