LinkedIn MCP Server & Personal AI Agent

An MCP server exposing 73 LinkedIn tools, plus an agent that ingests world news, clusters it, grounds claims against primary sources, writes the post, renders the card, and publishes, with an approval gate, a cost ceiling and a full action audit trail.

73 tools over stdio and Streamable HTTP, with an OAuth 2.1 layer for non-loopback MCP clients

Ingestion across 63 sources in 5 pillars (search, AI, developer, platforms, commerce) with feed-format detection, conditional requests and a robots-aware article fetcher

Grounding gate: every factual claim is tied to extracted facts with provenance, and a claim without a primary source is marked unconfirmed rather than published as fact

Model split by capability: DeepSeek for analysis and writing, OpenAI for embeddings and images, with an honest, named skip when a capability is unconfigured

Spend control at the model boundary: a usage ledger, a daily budget gate that refuses a call before it is made, and a per-image cost estimate

Schema discipline enforced by tests: every table single-writer, every column either written or the emptiness explained

750+ tests, including guards that fail the build if a source is claimed first-party without evidence