The Model Context Protocol is Anthropic's open standard for connecting AI agents to external tools and data sources via a structured client/server interface. This morning, a new release of https://bytecode.news provides an MCP endpoint.
The server is read-only and lives at https://api.bytecode.news/mcp. It exposes seven tools:
search_posts- full-text search across published postsget_post- fetch a single post by UUID, slug, or URLlist_taxonomy- get the managed set of categories and tagslist_factoids- paginated list of all factoids in the systemget_factoid- fetch a single factoid by selectorsearch_factoids- search factoids for termsfactoid_write_reference- returns command syntax and agent guidance for drafting new factoids
That last one is the interesting one. Factoids in nevet - the bot that connects to IRC, Slack, and Discord - are created via bot commands, not API mutations - !selector is information here, with optional attribute lines for URLs, tags, and cross-references. The MCP server can't write them, but it can tell an agent exactly how to draft one for human review. The factoid_write_reference endpoint returns examples, attribute keys, and an agentGuidance block that steers agents toward lean drafts: minimal tags, no speculative cross-references, leave the ambiguous bits for a human to fill in.
The read-only constraint is deliberate. Factoids are human-managed by design - they're tools the community uses to communicate with itself, and it's not something an API can really leverage well; an API will struggle with humor, with intent. nevet serves communities, and those communities should have input on what their information knowledge base looks like. The MCP surface is for agents that need to read content, answer questions about it, or prepare drafts that a human then commits. That division of labor is worth preserving.
If you want to connect to it, it works with any MCP-compatible client. Claude's connector configuration takes the URL directly, and any other connector should be able to work with it as well.