margrave MCP server

An MCP server over a shared store of agent skills (folders with a SKILL.md), run by margrave: the model describes its task in plain words, gets ranked skills, and loads only the one it needs — so a large skill library costs nothing until it is used. Read-only.

MCP on-demand rate limited read-only

Connect an MCP client

Remote streamable-HTTP endpoint — no auth, read-only: https://margrave.valksor.tools/mcp

Claude Code — CLI

claude mcp add --transport http margrave https://margrave.valksor.tools/mcp

Cursor · Cline — remote url (~/.cursor/mcp.json)

{
  "mcpServers": {
    "margrave": { "url": "https://margrave.valksor.tools/mcp" }
  }
}

VS Code — .vscode/mcp.json

{
  "servers": {
    "margrave": { "type": "http", "url": "https://margrave.valksor.tools/mcp" }
  }
}

Claude Desktop · Windsurf · other stdio clients — bridge via mcp-remote (claude_desktop_config.json)

{
  "mcpServers": {
    "margrave": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://margrave.valksor.tools/mcp"]
    }
  }
}

Claude Desktop only accepts command-based (stdio) servers in its config, so a bare url is rejected — the bridge above wraps the remote endpoint. On Pro/Team/Enterprise you can instead add it under Settings → Connectors → Add custom connector and paste https://margrave.valksor.tools/mcp.

Codex CLI — ~/.codex/config.toml

[mcp_servers.margrave]
command = "npx"
args = ["-y", "mcp-remote", "https://margrave.valksor.tools/mcp"]

Rule of thumb: clients with native streamable-HTTP/SSE support take the url directly (Claude Code, Cursor, Cline, VS Code); stdio-only desktop apps wrap it with npx -y mcp-remote https://margrave.valksor.tools/mcp.

Three tools

find_skills
task — the task in plain words; optional limit (default 5, 1–25).
Returns ranked skill ids with descriptions.
load_skill
id — an id from find_skills.
Returns the SKILL.md text, its source and the skill's file list.
read_skill_file
id, path — a path from that file list.
Returns one file; paths outside the skill are refused.

Example

Ask for skills, then load the best match. Each call is a JSON-RPC tools/call body posted to https://margrave.valksor.tools/mcp:

{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"find_skills","arguments":{"task":"review a React component"}}}
{"jsonrpc":"2.0","id":2,"method":"tools/call","params":{"name":"load_skill","arguments":{"id":"<id from find_skills>"}}}

The same first call as one line you can copy:

curl -X POST https://margrave.valksor.tools/mcp -H 'Content-Type: application/json' -H 'Accept: application/json, text/event-stream' -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"find_skills","arguments":{"task":"review a React component"}}}'

How it works

find_skills ranks the installed skills with BM25 over each skill's name, description, headings and body, weighting the name and description most. Skills load on demand: a search returns only ids and descriptions, and the full SKILL.md comes back only when the model calls load_skill. The skills live in a store on the server, not in your client; see what is installed now on /skills.