blade-mcp

An MCP server for the archive of the Ruby mailing lists (ruby-core, ruby-dev, ruby-list, ruby-talk, ruby-ext and ruby-math). It imports the original messages from the private blade-data-vault bucket into Postgres and serves read-only tools at POST /mcp over Streamable HTTP in stateless mode: search, get_message and get_thread over the mail, and search_matz and matz_timeline over what matz said about the design of Ruby, drawn from his mails and bugs.ruby-lang.org comments.

claude mcp add --transport http blade https://blade.ruby-lang.org/mcp --header "Authorization: Bearer $BLADE_MCP_TOKEN"

Responses carry the sender's name with the domain masked (matz@...), the date, the subject, the decoded body and attachments. To, Cc, Reply-To, Return-Path, Received and Message-ID are never returned. Redmine notification mails keep only the subject and the issue number, which the bugs.ruby-lang.org MCP server takes.

The skill in skills/blade shows Claude how to use these tools in the core team's daily work, from finding past discussions and judging who decides a question and how matz would see a proposal, to drafting an agenda item for the developers' meeting. Install it as a personal skill:

mkdir -p ~/.claude/skills/blade && curl -fsSL https://raw.githubusercontent.com/ruby/blade-mcp/main/skills/blade/SKILL.md -o ~/.claude/skills/blade/SKILL.md

Configuration

Variable	Purpose
DATABASE_URL	Postgres with the vector extension
BLADE_MCP_TOKEN	Passphrase expected in Authorization: Bearer
AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY	Read access to blade-data-vault, for the import commands
EMBEDDING_URL, EMBEDDING_KEY	Heroku Managed Inference, cohere-embed-v4 unless EMBEDDING_MODEL_ID says otherwise
RERANK_URL, RERANK_KEY	Heroku Managed Inference, cohere-rerank-3-5 unless RERANK_MODEL_ID says otherwise
INFERENCE_URL, INFERENCE_KEY	Heroku Managed Inference, claude-opus-4-8 unless INFERENCE_MODEL_ID says otherwise, for extract
BUGS_DATABASE_URL	The bugs.ruby-lang.org database, read once by import-redmine
GITHUB_TOKEN	Optional token for the GitHub API calls that sync ruby/dev-meeting-log, which are limited to 60 an hour per address without one

Without the inference variables, search ranks by the full-text index alone.

Commands

bin/blade-mcp migrate
bin/blade-mcp import --lists ruby-dev --limit 2000
bin/blade-mcp import-redmine
bin/blade-mcp import-meetings
bin/blade-mcp update
bin/blade-mcp embed --lists ruby-talk
bin/blade-mcp extract --sources redmine --limit 100

import reads whole lists, update takes only messages numbered above the last one stored, embeds them, fetches bugs.ruby-lang.org comments by or about matz through the REST API and the meeting notes changed in ruby/dev-meeting-log, extracts statements from what is not yet read and embeds those statements, which is what Heroku Scheduler runs daily. import-redmine fills those comments in from the Redmine database once, and import-meetings the notes of the developers' meetings from a tarball of the repository, split into agenda items at their headings; both have to run before the first update. embed covers every list but ruby-talk by default and skips Redmine notifications. extract has the chat model read matz's mails, with the post each one replies to, those comments and the agenda items, and records what he said about Ruby's design as statements: decisions, opinions, conditions and principles, each with its reason and the features it concerns. Every mail, comment and agenda item is read once.

Development

Schema changes are Sequel migrations in db/migrate, named with a timestamp and written as raw SQL with run, and migrate applies the ones not yet recorded, which Heroku does on every release. Tests drop and recreate the tables in TEST_DATABASE_URL, which defaults to postgres://postgres@localhost/blade_mcp_test and needs pgvector.

createdb blade_mcp_test
bundle exec rake test