← Back to Insights
ORIGINAL 2026-10-07 · AI Development & Integration

Search, Handshake, Paste — An Agent Skill for the Influzer MCP Catalog

Discovery can search. Probe can handshake. Snippets can paste. We shipped the SKILL.md that teaches agents that loop — and we still will not call it a safe-to-install badge.

MCP Model Context Protocol Agent Skills Cursor Claude ChatGPT ACP Directory Governance
Rasmus Mencke
Technology & product leader · 3 min read
Connecting Influzer MCP Discovery so an agent can search, handshake, and paste client config

Directories taught agents how to search. Clients taught humans how to paste. Nobody taught the agent the sequence in between, so it hallucinated an npx line and called it setup.

Influzer now ships that sequence as an agent skill: search, handshake, paste. Discovery looks up the listing. Probe talks initialize then tools/list. Install snippets return Cursor JSON, Claude connector steps, ChatGPT custom connector, Claude Code, and ACP mcpServers — in the same get_mcp_server result. safety_badge is still null.

Skill page: /mcp/discovery/skill · Raw: /skills/influzer-mcp/SKILL.md

The hole after snippets

Last week we published per-listing install snippets so a README or a docs site can embed the same paste targets. That is the human path.

The agent path was still broken. Discovery could search. Probe could handshake. The model still had to guess which JSON went in .cursor/mcp.json, which URL belonged in a Claude connector, and whether ChatGPT could even reach the process. Stdio listings got sold as “works in Claude.” They do not.

A skill is the smallest object that fixes that. It is markdown with a job. It is not a new MCP server, not a WebMCP page, not an MCP App, and not ACP.

What we shipped

  1. SKILL.md at /skills/influzer-mcp/SKILL.md — search → handshake → paste, plus the protocol split.
  2. In-band snippets on Discovery get_mcp_server. The tool result now includes install.clients. The agent does not need a second fetch to know what to paste.
  3. A public install path — drop the file into .cursor/skills/influzer-mcp/ or .claude/skills/influzer-mcp/. Index: /.well-known/agent-skills/index.json.
mkdir -p .cursor/skills/influzer-mcp && curl -fsSL -o .cursor/skills/influzer-mcp/SKILL.md \
  https://www.influzer.ai/skills/influzer-mcp/SKILL.md

The loop, as the skill writes it

  1. Search with Discovery (https://www.influzer.ai/mcp/discovery) or open the listing.
  2. Read install on get_mcp_server. If stdio_only, do not give Claude chat or ChatGPT a localhost URL.
  3. If there is an HTTPS remote, GET /api/v1/probe. live_ok or auth_required is a handshake. unreachable is a ticket. Probe is not a SAFE badge.
  4. Paste the snippet for this client. A human still allowlists writes.

Discovery remains read-only. It does not install. That split is still policy before plugins.

Four sockets, still four

The skill is explicit because the hallway is not:

If a vendor cannot point at one of those, they are selling a slogan. The skill will not “search ACP” or treat a site tool as a remote connector.

What this is not

It is not NVIDIA SkillSpector. Scan cloneable skill source when you have source — we already wrote that split. A remote HTTPS MCP has no zip to lint. Handshake it.

It is not an auto-connect. Ranked search is not an allowlist. A live tools/list is not “safe to give the agent GitHub writes.”

It is not a new catalog. Same Influzer listings, same honesty labels, same refusal to print SAFE.

Quick answers

Do I need Discovery connected for the skill to work?

It is the intended path. Without it, the skill still points at /api/v1/install/<slug> and the probe URL. Connect Discovery once: setup guide.

Will the skill install MCP servers by itself?

No. It pastes. You still own .cursor/mcp.json and the Claude connector list.

Can ChatGPT use a stdio listing?

No. The snippets already say that. The skill repeats it.

Is this an MCP App?

No. There is no host iframe. It is a skill file.

Final thought

Search without handshake is a blog. Handshake without a paste target is a JSON souvenir. Paste without a protocol split is how localhost lands in ChatGPT.

Search. Handshake. Paste. Never badge it SAFE.

Start with the skill. Keep Discovery read-only. Keep writes on a list you can name.

GET PRACTICAL AI PLAYBOOKS WEEKLY

One clear email each Thursday

Actionable frameworks on AI execution, agents, and MCP. Join 4,200+ builders.

✓ You're in — first briefing Thursday.
Written by Rasmus Mencke
Technology and product leader sharing practical insights on AI, leadership, and scaling teams.

Leave a comment

Be the first to share your thoughts.

Related insights

2026-10-02
ACP vs MCP — The Editor Layer Is Not the Tool Layer
Agent Client Protocol is how Zed and JetBrains host a coding agent. MCP is how that agent reaches tools. They share JSON-RPC. They do not share an allowlist. Spell the words before you pick a winner.
2026-10-02
MCP Apps Are Not WebMCP
SEP-1865 puts a sandboxed widget inside Claude and ChatGPT. WebMCP puts tools on the live page. Classic MCP is still just JSON. Pick the surface before Toronto sells you all three as one demo.
2026-10-01
Pack an Allowlist for MCP Dev Summit
MCP Dev Summit opens in Toronto on 5–6 October. The hallway will sell you plugins. Bring a map of the remotes you already run — and a one-page allowlist — or you will come home with stickers.