Overview

Toolbox is five small tools that each do one job an agent builder keeps needing. Every tool runs three ways: as a form on the homepage, as POST /api/<tool> with a JSON body, and as a tool on the MCP server at /mcp. None of them call a language model. The code is deterministic, so the same input gives the same answer, and nothing here spends your API key or ours.

Free tier and Pro keys

No signup for the free tier. Each tool has its own daily cap per caller IP, listed below, and everyone gets 30 requests a minute. Pro is $5 a month: no daily caps, 300 requests a minute, larger inputs. Start it from the homepage; the key is emailed to you and shown once on the confirmation page. Pass it as Authorization: Bearer tb_... on any route, including /mcp. Manage the subscription at /billing.

Tells

Finds the patterns readers use to spot model-written text and returns a cleaned copy. Checks: em-dashes, "not just X but Y" and "it's not about X, it's about Y" framings, three-item lists, stock vocabulary (delve, seamlessly, leverage, robust, game-changer and about forty more), hedge words, curly quotes, hidden Unicode (zero-width, bidi, TAG block), conversational openers, "whether you're" audience splits, and rhetorical questions answered by the next line. The score weights the tells by how reliably each one gets a piece flagged. Cleaning replaces em-dashes with a comma or a period depending on what sits either side, straightens quotes, and strips invisible characters. It does not rewrite sentences.

curl -X POST https://toolbox.allthepossibles.com/api/tells \
  -H "content-type: application/json" \
  -d '{"text": "It is not just fast, it is a game-changer."}'

Returns score (0 to 100), verdict (clean, light, heavy), counts by kind, tells with positions and excerpts, and cleaned. Free: 20 runs a day, 3,000 characters per run. Pro: 200,000 characters. MCP tool: check_tells.

Manifest cost

Every MCP client loads a server's full tool list at session start, so a chatty server charges its users on every session before the first call. This tool does the real handshake (initialize, session id, initialized notification, tools/list) against a Streamable HTTP endpoint, or takes a pasted tools array, and estimates tokens per tool at JSON characters divided by 3.8, with base64 runs counted at 1.5 characters a token. It names the three most expensive tools and flags descriptions over 600 characters, base64 blobs, schemas over 1,500 characters, descriptions that repeat their own argument text, and output schemas.

curl -X POST https://toolbox.allthepossibles.com/api/manifest-cost \
  -H "content-type: application/json" \
  -d '{"url": "https://preflight.allthepossibles.com/mcp"}'

The number is an estimate; real tokenizers differ by model. Free: 10 runs a day. MCP tool: manifest_cost.

Agent file lint

Lints a CLAUDE.md, AGENTS.md or MEMORY.md. Reports the line count against the 200-line cap Claude Code applies to the memory index (lines past it are never read), estimated tokens reloaded per turn, duplicate lines, lines over 200 characters, pairs of rules that give opposite instructions about the same thing, empty headings, leftover TODO and FIXME markers, em-dashes, and hidden characters.

curl -X POST https://toolbox.allthepossibles.com/api/agent-md-lint \
  -H "content-type: application/json" \
  -d '{"text": "# Rules\n- Always run tests.\n- Never run tests on main.", "filename": "CLAUDE.md"}'

Pass filename so the line cap is scored correctly: it is high severity for MEMORY.md and medium for everything else. Free: 20 runs a day, 3,000 characters per run. MCP tool: lint_agent_md.

Install snippets

Give it an MCP endpoint and it returns the snippet for Claude Code (claude mcp add), Claude Desktop, Cursor, Codex CLI, VS Code, and a two-step curl handshake, with an optional auth header threaded through each. It also probes the endpoint once and tells you whether it answered initialize and tools/list, how many tools it has, and how long the handshake took, so a broken config line and a broken server stop looking the same.

curl -X POST https://toolbox.allthepossibles.com/api/mcp-install \
  -H "content-type: application/json" \
  -d '{"url": "https://preflight.allthepossibles.com/mcp", "name": "preflight"}'

Free: 30 runs a day. MCP tool: mcp_install_snippets.

AI crawlers

Reads a site's robots.txt the way the crawlers do (longest matching user-agent group wins, then the wildcard) and reports allowed, blocked or partial for nineteen AI agents: GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, Claude-User, Google-Extended, PerplexityBot, Bytespider, CCBot, Amazonbot, Applebot-Extended, cohere-ai, Meta-ExternalAgent and the rest. Tells you whether llms.txt exists, and drafts one from the page title, description and the first twenty sitemap URLs.

curl -X POST https://toolbox.allthepossibles.com/api/ai-crawlers \
  -H "content-type: application/json" \
  -d '{"url": "https://allthepossibles.com"}'

Free: 10 runs a day. MCP tool: check_ai_crawlers.

MCP server

claude mcp add --transport http toolbox https://toolbox.allthepossibles.com/mcp

{
  "mcpServers": {
    "toolbox": { "type": "http", "url": "https://toolbox.allthepossibles.com/mcp" }
  }
}

Five tools, all annotated read-only: check_tells, manifest_cost, lint_agent_md, mcp_install_snippets, check_ai_crawlers. Arguments match the REST bodies. A Pro key as a Bearer token on the MCP connection lifts the caps there too.

Limits

Everyone has a burst cap: 30 requests a minute free, 300 on Pro. A 429 carries the reason in error. Caps reset at midnight UTC.

Support

Bugs, a tell the checker misses, a crawler that should be on the list: support@allthepossibles.com. Billing: /billing. Usage this month: GET /api/usage with your key.