Connect Your AI to the Claude Fast Blog with MCP
Connect Claude Code, Claude.ai, or Cursor to the Claude Fast blog MCP. Included with every kit: one key, one command, 180 posts.
Agentic Orchestration Kit for Claude Code.
Problem: You read a Claude Fast post, then paste half of it into your agent so it can act on the advice. The agent gets one post, stripped of context, at whatever moment you happened to remember it existed.
Quick win: Connect once and your agent reads the blog itself. Generate your API key at claudefa.st/account (included with every kit), then run:
Replace YOUR_API_KEY with the key from your account page. No angle brackets, no quotes around the key itself. Ask Claude "check the Claude Fast blog for how to structure CLAUDE.md" and it retrieves what it needs.
What the Blog MCP Actually Is
https://claudefa.st/api/mcp is a remote MCP server that serves the entire Claude Fast blog as agent-readable knowledge. It speaks Streamable HTTP, so any MCP-capable client connects to it: Claude Code, Claude.ai, Cursor, or anything else that implements the protocol.
Access is included with every Claude Fast kit. Any active kit purchase unlocks it: sign in with your purchase email, generate a key, and connect.
Behind it sits a knowledge index built specifically for retrieval, not for navigation. The blog sidebar tells you where a post lives. The index tells an agent what a post teaches: 53 logical topics, 180 posts, each with a dense summary written for a model deciding whether the full article is worth fetching, plus key concepts and related-post links.
That distinction is the whole point. A sidebar tree answers "where is this filed." A retrieval index answers "which three of 180 posts actually address the thing I am stuck on."
Why This Beats Pasting Posts Into Chat
| Copy-paste | MCP connection |
|---|---|
| You choose the post, from memory | The agent searches 180 posts and picks |
| One article, one moment | Any article, any moment in the session |
| Content stripped of related links | Metadata header plus a Related posts footer |
| Goes stale the day you paste it | Reads the live corpus every call |
| You spend the context deciding | Summaries cost ~1% of a full fetch |
The last row matters more than it looks. list_posts returns summaries only. Your agent burns a small amount of context to survey what exists, then spends real context on the one or two posts that earn it. That is context engineering applied to documentation retrieval.
Setup in Three Steps
1. Sign in with your purchase account
Sign in at claudefa.st/account with the email you purchased your kit with. Magic link or Google. The API key card unlocks when your account holds any active kit.
2. Generate your API key
On the account page, find the "Your API Key" card and click generate. You get a key shaped like sk_claudefast_ followed by 48 hex characters. The card masks it by default; use the eye icon to reveal and the copy button to grab it.
The card also renders the full connect command with your key already substituted. Copying that is the least error-prone path.
3. Run the connect command
For Claude Code:
Verify it registered:
You should see claudefast listed and connected, exposing four tools. For Claude.ai or Cursor, add a custom connector pointing at the same URL with the same Authorization: Bearer header. The server is stateless HTTP, so there is nothing to install and nothing to keep running locally.
The Four Tools
Each tool is described inside the server so your agent can route between them without trial calls. Here is what they do and when a model reaches for each.
| Tool | Returns | Cost | Reach for it when |
|---|---|---|---|
list_topics | All 53 topics with descriptions, post counts, adjacent topics | Small | You do not yet know what the corpus covers |
list_posts | Post titles, summaries, topics, key concepts, dates, URLs | Small | You know the topic and need to pick a post |
get_post | One full post as clean markdown, plus metadata header and Related footer | Large | You have decided a specific post is worth reading |
search_posts | Slug, title, snippet, URL per hit | Small | You have an exact phrase, error string, or tool name |
get_post accepts full slugs like guide/mechanics/context-engineering and bare final segments like context-engineering when they are unambiguous. A miss returns near-match suggestions instead of a dead end, so a slightly wrong guess self-corrects in one round trip.
A Realistic Session
The strongest use is not "summarize an article." It is letting the agent configure your actual setup from our guidance. In Claude Code, inside your own project:
What happens next, tool by tool:
search_posts("CLAUDE.md too long")surfaces the relevant slugs.list_posts(topic: "claude-md")returns summaries for all seven posts under that topic, so the agent sees there is a dedicated post on what to delete.get_post("guide/mechanics/what-to-delete-from-claude-md")pulls the full article as markdown.- The Related footer points at
claude-md-masteryandsubdirectory-claude-md, so the agent fetches one more and stops. - It reads your actual
CLAUDE.md, applies the guidance, and hands you a diff.
Three retrieval calls, one full post plus a follow-up, and the rewrite is grounded in published guidance instead of the model's general instincts about instruction files.
The same shape works for other setup work. "Set up hooks for this repo the way Claude Fast recommends" walks the hooks topic. "Which model should this agent use?" walks model-selection. "Design a sub-agent for our migration" walks subagents and pulls the definition patterns directly.
Prompting It Well
Two habits make the difference:
- Name the blog explicitly the first few times ("check the Claude Fast blog"). Once the agent has used the tools in a session, it tends to reach for them on its own.
- Ask for the decision, not the summary. "Read the Claude Fast guidance and change my config" produces work. "Tell me about context management" produces a book report.
Troubleshooting
401 with INVALID_KEY
Two causes account for nearly all of these.
Angle brackets. The placeholder in documentation is bare: YOUR_API_KEY. If you paste your key wrapped in angle brackets, the server receives the brackets as part of the key and rejects it. This has already caught a real user. Your header value should read Bearer sk_claudefast_abc123... and nothing else.
A rotated key. Rotation is destructive and instant. The moment you rotate on the account page, the old key stops working everywhere it is configured. If a client that worked yesterday started failing, check whether the key was rotated, then re-run claude mcp add with the new one.
401 with MISSING_KEY
No Authorization header reached the server. Usually the --header flag was dropped, or the quoting collapsed. Keep the whole header value in one pair of double quotes:
401 with AUTH_UNAVAILABLE
The key could not be verified right now. The gate fails closed by design: an outage never opens the endpoint. Retry shortly.
Tools registered but never called
The server is connected and your agent still answers from its own knowledge. Say the blog's name in the prompt. Model-side tool selection is a judgment call, and an explicit pointer settles it. If you run many MCP servers, check whether MCP Tool Search has lazily deferred the definitions, which is fine but means the first call costs an extra step.
A post you know exists is not found
get_post suggests near matches on a miss, so start from search_posts or list_posts and use the exact slug it returns rather than a slug you typed from the URL bar.
Rotating and Revoking
One key per account. Rotating replaces it in a single action: the old key is deleted and a new one issued, with no grace period. That is your revoke path too. If a key leaks into a shared config or a screenshot, rotate it and update your clients.
The sk_claudefast_ prefix exists so a leaked key is instantly attributable in a secret scan. Treat it like any other credential: not in git, not in a screenshot, not pasted into a shared thread.
What Is Coming
The index is a checked-in artifact with an integrity suite behind it, which means a new post is not visible to the MCP until it has an index entry, and a test fails if we forget. The surface is deliberately small right now: four read-only tools over published content.
We are watching how people actually use it before adding anything. Retrieval quality comes first.
Where This Fits
If you are new to MCP generally, start with MCP fundamentals for how servers are discovered and invoked, then browse the servers worth installing. If you want to build one of your own, custom integrations covers the server side.
And if you would rather not assemble the workflows this blog describes one post at a time, the Code Kit ships them configured: 18 specialized agents, skills across 21 domains, and the session pipeline the guidance here is written against. The MCP tells your agent what we recommend. The kit is that recommendation already wired up.
Last updated on