Model Context Protocol
Well-crafted MCP servers.
mcpwright (noun) — a maker of MCP servers.
Each one brings a real-world data source into your agent through a tight,
well-documented tool surface: typed structured output, clear descriptions
the model can reason about, and read-only by default.
Servers
● liveedgar-mcp
SEC EDGAR inside your agent — resolve any company (public or private), browse Reg CF / Reg D / Reg A raises, screen the actual deal economics, and read insider trades. 11 tools over public SEC data.
● livecensus-mcp
U.S. Census data by ZIP — income, demographics, housing, and education from the Census Bureau's ACS, plus place → ZIP reverse lookup. 8 tools, bulk-downloaded once into a local store, then served instantly offline.
● livesoi-mcp
IRS income & tax by ZIP — the real AGI distribution, tax, credits, and deductions from filed returns (IRS Statistics of Income), not survey estimates. 10 tools, no API key, bulk-loaded once then served offline.
● livefred-mcp
FRED economic data, with a memory — 800,000+ time series plus ALFRED vintages: any series as it was known on a past date, and any data point's full revision history. 9 tools, revision-aware caching, bring your free key.
More on the way.
How these are built
- Official SDK. Python, on Anthropic's official
mcpSDK. - Typed, structured output. Tools return Pydantic models, so agents get data, not just text.
- Read-only by default. Annotated and safe to run, with clear, actionable errors.
- Easy to run.
uvx <server>-mcp, with copy-paste config for Claude Desktop & Code. - Tested & CI-gated. ruff + mypy + a real test suite on every change.
The thinking behind these → What makes a good MCP tool surface for an LLM