MCP server

Let your agent ask which skill to use — then hand it over.

An open, read-only MCP server over all 50 book-based skills and 12 guided journeys. It routes a task to the right framework, tells your agent which ones to skip and why, and serves the real SKILL.md in-session. No install, no account, no auth.

POST https://skills.wondel.ai/mcp
The round trip

One task in, the right method out

  1. 01 your task “our landing page gets traffic but nobody signs up”
  2. 02 recommend_skills which frameworks to load — and which to skip, whenever a skill’s own “not for” line rules it out
  3. 03 load_skill the upstream SKILL.md, ready to follow in this session
  4. 04 your agent works applying the method instead of improvising one

Nothing is installed and nothing is cached on your side — steps 02 and 03 are two ordinary tool calls, and your agent decides whether to make them.

Connect it

Streamable HTTP, one URL, nothing to configure. In Claude Code:

claude mcp add --transport http wondel-skills https://skills.wondel.ai/mcp

In any client that takes an mcpServers block — Cursor, Codex, OpenClaw and the rest: 

{
  "mcpServers": {
    "wondel-skills": {
      "type": "http",
      "url": "https://skills.wondel.ai/mcp"
    }
  }
}

5 tools

Every one is read-only and annotated as such, so clients can call them without asking you first. Start with recommend_skills when you have a task and search_skills when you have a keyword.

  • search_skills (query, kind?, category?, limit?)

    Search 50 book-based agent skills and 12 guided journeys by keyword or phrase. Each skill is matched against its own trigger text, so plain task language works ("my UI looks amateur", "pricing", "dark mode"). Returns compact hits, each quoting the line that matched. Use recommend_skills instead when you have a task rather than a keyword.

  • recommend_skills (task, stack?, phase?, max?)

    Route a task to the right skills. Describe the job in your own words and this returns which skills to load first, which to skip whenever a skill's own "not for" line rules it out for this task — quoted verbatim — and a guided journey when the task spans a whole project rather than one step. This is the tool to reach for first.

  • get_skill (slug, sections?)

    Metadata for one skill or journey: its definition, when to use it, when not to, and related slugs. Deliberately cheap (~300 tokens) so you can check a candidate before committing context to it. Pass sections to widen, or sections:["all"] for the whole page. This does not return the skill instructions — load_skill does.

  • load_skill (slug, part?)

    The skill's actual instructions — the upstream SKILL.md, ready to follow in this session with nothing installed. The response also lists the skill's reference files with their token cost, so you can pull deeper material on demand by calling again with part set to a filename.

  • get_journey (slug? | goal?)

    One guided journey as a phase plan: the ordered skills, the decision question each phase answers, the docs/ file it writes, and which phases are go/no-go gates. Pass slug for a known journey, or goal to have one picked.

Each also takes a context string — one line on the task you are solving. It never reaches the tool, it is stripped before the handler runs; it is there so the usage numbers say what agents came here to do rather than only which tool they pressed. Leave it out and the call still works. What we record per call — tool, arguments, a truncated result, duration, your client's user agent — goes to our own PostHog on EU servers. No account, no auth, no cookie, and nothing that identifies you beyond that.

Resources

For clients that attach context directly instead of calling tools. The skill:// template is enumerable and slug-completable, so all 62 show up by name.

  • wondel://catalog

    An index of all 62 skills and journeys in one ~15 KB JSON read — slug, name, one-liner, category. Use it to see what exists; get_skill returns when-to-use detail for the few you care about.

  • wondel://journeys

    The 12 guided journeys with their full phase plans, artifacts and gates.

  • skill://{slug}

    One skill's upstream SKILL.md. Enumerable and slug-completable, so clients can offer all 62 by name.

  • skill://{slug}/references/{file}

    A skill's deep-dive reference — the progressive-disclosure layer under SKILL.md.

14 prompts

In clients that surface MCP prompts as slash commands, each guided journey becomes one — it walks you phase by phase, asks you the decision question at each step, and writes the answers to your docs/ folder.

  • /create-business Raw idea → validated, positioned, priced business
  • /create-website Blank page → high-converting website
  • /create-app Idea → validated, well-architected app
  • /improve-business Diagnose and fix a plateaued business
  • /improve-website Diagnose, fix, and prove website improvements
  • /improve-app Product, UX, copy, and in-app persuasion — fix what users feel
  • /grow-business Build a repeatable growth engine
  • /grow-website Convert, capture, and compound website traffic
  • /grow-app Retention-first app growth
  • /improve-code-quality Vibe-coded prototype → production-ready code
  • /remove-technical-debt Pay down debt in place without stopping shipping
  • /design-code-architecture Deliberate architecture for a new system
  • /pick-skill (task)

    Hands the model the full catalog and asks it to choose and load the right skill for a task you describe.

  • /audit-with (slug, target)

    Loads a named skill and applies it as an audit to a file, directory or URL.

We argue skills usually beat MCP. So why this?

Our own explainer on skills versus MCP makes the case that a skill is instruction and MCP is connection, and that reaching for a server when your agent can already read the thing adds a service to operate without adding any access. We still think that.

This server does not contradict it, because it is not connecting your agent to a system — it is solving the two problems that a library of 62 frameworks has and a single skill does not. Which one? Routing is a real question when the wrong choice costs you a context window; recommend_skills answers it from each skill's own hand-written “not for” list, which is the part no keyword search can reproduce. How do I try one? Reading a skill before installing it used to mean browsing GitHub; now it is one tool call.

For everyday work, install what you use — npx skills add wondelai/skills --all --global — and skip the network entirely. Everything here is MIT-licensed and readable at github.com/wondelai/skills, built on the open agentskills.io standard.

FAQ

Frequently asked questions

Do I still need this if I have already installed the skills?

Mostly no, and that is the honest answer. Installed skills live on disk, load with no network hop, and your agent picks them up by itself. The server earns its place in three situations: an agent that has nothing installed and needs a framework now; choosing between 62 entries, which is what recommend_skills is for; and reading a skill's deep-dive reference files on demand instead of carrying them in context.

What does connecting cost me in context?

About 2,600 tokens for the tool definitions, paid once per session — that is the real price of connecting, and it is why there are five tools rather than twenty-five. After that the surface is deliberately cheap: a search hit is ~70 tokens as text and ~120 if your client reads the structured lane, and get_skill ~400. Only load_skill spends real context (~4,500 tokens for an average SKILL.md), because that is the call that returns instructions to follow — and it returns them once, in one lane, rather than paying for the same document twice.

Is there authentication or a rate limit?

Neither. Every tool is read-only, there is no state and no account, so the endpoint is open and CORS is unrestricted — browser-hosted MCP clients work too. Tool calls are answered live; it is the reads behind them — the upstream SKILL.md files and the repo file tree — that are cached at the Cloudflare edge for an hour, which is why a warm load_skill returns in about 60ms. If you want a guarantee rather than a promise, install the skills instead: npx skills add wondelai/skills --all --global.

Which clients does it work with?

Anything that speaks streamable HTTP MCP — Claude, Claude Code, Claude Cowork, Codex, Cursor, OpenClaw, Hermes Agent and others. The server implements the stateless 2026-07-28 revision and still answers older stateless clients on the same URL, so there is no separate legacy endpoint and no SSE session to keep alive.

Work with us

We build the skills you already use. Now we’ll build yours.

Custom skills · Subagents · MCP integrations — shipped to production, not demoed.

See what we build Book a 15-min call

Sprints from $3K · shipped to production, or you don’t pay the final milestone.