Bumblebot is a self-hosted Telegram + Claude Code chassis with persistent memory and a nightly dreaming loop. Bring your own Postgres, bring your own keys, run it on your own box.
git clone https://github.com/mattsmith/bumblebot && cd bumblebot && ./installFive prompts, ~15 minutes from clone to first message. Needs Node 20+, Postgres (free Supabase works), an Anthropic API key, and a Telegram bot.
Allowlist-gated. Replies inline, mirrors into forum topics, ignores everyone you didn't invite.
Drop a <<DISPATCH>> marker into a message. Bumblebot spawns Claude Code in a fresh git worktree, opens a PR, replies with the link.
Hybrid pgvector + BM25 recall. Every meaningful turn is captured with embeddings, full-text, agent namespace, and a significance score.
Every six hours, the bot dedupes near-duplicates, strengthens repeated facts, and synthesizes clusters into compiled memories.
Indie hackers and solo builders who already run their own infrastructure, are comfortable with a CLI, and want a personal assistant without giving Anthropic, OpenAI, or a SaaS vendor their entire memory graph.
.env is gitignored, a pre-commit hook blocks accidental commits, and there is no PUBLIC_MODE flag.
Running publicly requires forking.
<<DISPATCH>> → git worktree → Claude Code → PR flowbrain, graph_entities, graph_edges, sessions (migrations included)bumblebot doctor — audits your install for security regressionsbumblebot start | stop | recall | capture | dream-nowv1 spec is published; implementation is in progress. Track progress in the repo or read the full spec.