For agents and the people who build them

Everything here is readable by a machine

The site publishes its content as markdown, its actions as an OpenAPI spec, and its structure as a plain-text index. No scraping, no HTML parsing, no guessing.

Start here

Discovery

Four URLs describe the whole site. Fetch any of them without a key.

Site index: /llms.txt — Every public page, one line each, with a description.
Full content: /llms-full.txt — Prose: services, pricing, how we work, FAQ.
Actions: /openapi.json — OpenAPI 3.1 for every action an agent can take.
WebMCP: /mcp-actions.json — The same actions as WebMCP task descriptors.
Sitemap: /sitemap.xml — Canonical URLs with real lastModified dates.
Feed: /rss.xml — Full blog feed.
MCP server: /mcp — Streamable HTTP. Every action above, plus blog search and markdown, as MCP tools.

Content

Any post, as markdown

Append .md to a post URL, or ask for the canonical URL with an Accept header. Both return YAML frontmatter plus the body — no nav, no chrome, no cookie banner.

markdown

# the .md suffix
curl https://www.promptmetrics.dev/blog/ai-agents-in-production-problems.md

# or negotiate on the canonical URL
curl -H 'Accept: text/markdown' \
  https://www.promptmetrics.dev/blog/ai-agents-in-production-problems

# an index of every post
curl https://www.promptmetrics.dev/blog.md

A markdown request that misses returns a real 404 with a markdown body listing where to go instead, not an HTML error page.

Actions

Things an agent can do

Unauthenticated JSON endpoints. /v1 is canonical; the unversioned paths are permanent aliases kept for existing callers.

actions

# read the contract first
curl https://www.promptmetrics.dev/openapi.json

# subscribe to Field Notes
curl -X POST https://www.promptmetrics.dev/v1/actions/subscribe-blog \
  -H 'Content-Type: application/json' \
  -d '{"email":"[email protected]"}'

Every 4xx and 5xx returns { error: { code, message } } with a code from a published enum. Branch on the code, never on the message.

Read the OpenAPI spec

MCP

Or connect it as a server

The same actions, plus blog search and post retrieval, over Streamable HTTP. No auth, no session — point any MCP client at the endpoint.

mcp

# Claude Code
claude mcp add --transport http promptmetrics https://www.promptmetrics.dev/mcp

# or any MCP client, pointed at
https://www.promptmetrics.dev/mcp

Tools that submit a request on someone's behalf do exactly that. Confirm with the person before calling one.

Limits

What to expect

Auth: None. The action endpoints are public.
Versioning: In the path. Breaking changes ship under a new prefix, never in place.
Rate limit: 20 requests per 60 seconds per IP. Responses carry RateLimit-Policy when the limit is being enforced; a 429 carries Retry-After.
Errors: Typed codes: invalid_body, missing_field, invalid_email, unknown_segment, rate_limited, not_configured, upstream_error, internal_error.
Crawlers: AI crawlers are welcome. See /robots.txt — training is reserved via Content-Signal, retrieval is not.

Open source

The code we publish

Skills are folders of markdown, MIT licensed. There is no CLI and no npm package to install — you clone a repo and copy what you want.

HubSpot Plugin: Natural-language HubSpot CRM admin, with a human approving every write /plugin install hubspot@hubspot-claude
HubSpot MCP Server: A standalone MCP server for HubSpot CRM with a write-safety layer /plugin install hubspot-mcp@promptmetrics-hubspot-mcp
The Human Layer: Rehearse the hard conversation before you have it /plugin install promptmetrics-coach@promptmetrics
The Habit Layer: One sharp question before Claude runs a real task /plugin install the-habit-layer@habit-layer
Video Skill: Plan a Kurzgesagt-style explainer, ready to hand to a renderer /plugin install promptmetrics-video@promptmetrics-video
pm-blog: AI Blog Writing & SEO Optimization Skill for Claude Code /plugin marketplace add promptmetrics/pm-blog

All open source