← All projects
Active

8 Custom MCP Servers

Tooling for MCP-capable clients. Email, YouTube, audio analysis, crypto trading, job hunting, peer networking -- all as local MCP servers with zero cloud dependencies.

8
servers
80+
tools total
0
cloud dependencies
local
all self-hosted

What they are

MCP (Model Context Protocol) lets AI models call external tools natively. Instead of prompt-engineering an LLM to output JSON that some wrapper parses, MCP defines a standard interface: the model sees available tools, their parameters, and their return types. It calls them like functions. These 8 servers give MCP-capable clients access to everything from sending emails to analyzing audio to trading crypto -- all running on local hardware with no external service dependencies.

The Servers

PeerMesh · 21 tools
Multi-agent communication for MCP clients. Direct messages, broadcast channels, shared memory with FTS5, task delegation, stable identities, and cross-machine coordination.
Audio · 19 tools
Music analysis pipeline. Chord detection, stem separation, genre classification, BPM detection, melody extraction, lyric transcription.
Grayson · 13 tools
Job outreach pipeline. Lead management, email verification (MX + SMTP probe), tailored drafting, SMTP send with resume, reply tracking.
Goldmine · 9 tools
Crypto trading on Base network. Market scanning, coin analysis, paper/live trade execution, portfolio monitoring, signal detection.
YC/HN · 9 tools
Y Combinator and Hacker News intelligence. Company lookup, job scanning, thread monitoring, application tracking.
YouTube · 7 tools
Video intelligence. Search, metadata extraction, transcript pull, comment analysis, playlist enumeration, audio download, frame extraction.
DevMail · 7 tools
Email operations. Send, read, search, reply, folder management, count. IMAP + SMTP through Hostinger. Portfolio domain sender identity.
Blackreach · 3 tools
Autonomous web agent. General browsing, targeted scraping, and search — all through a stealth Playwright backend with DOM compression.

Why MCP

Before MCP, giving an AI tools meant writing custom JSON parsers, building wrapper scripts, and hoping the model's output matched your expected format. MCP standardizes this. The model sees a typed tool interface -- name, description, parameter schema, return type. It calls the tool. The server executes it. The result comes back structured.

The practical effect: an MCP-capable client can send an email, analyze a song, check crypto prices, and message another agent -- all in the same conversation, using native tool calls. No prompt engineering. No output parsing. No wrappers. 

claude > "Send my resume to that lead and check if any peers are online"

  Tool call: grayson_send(lead_id=12)
    Result: Email sent to jane@example.com [verified]

  Tool call: list_peers(scope="machine")
    Result: [{id: "agent-b", status: "online",
              summary: "Running test suite"}]

  Tool call: send_message(to="agent-b", body="Tests done?")
    Result: Message delivered

Built with Python (FastMCP) and TypeScript (Bun). All servers run as background processes on a single Linux machine. Each one is independently deployable and testable.

← Grayson Home →