I have been playing around with multi-agent orchestration, it have mostly been for my learnings but all code are opensource.
GitHub - Taure/banto: Multi-agent repo concierge for the BEAM - index your repos into shared semantic memory and ask across them via MCP. The gakudan ecosystem showcase. · GitHub this is the showcase repo.
banto (番頭, the head clerk who runs the shop) points a swarm of agents at your repositories. It indexes them into one semantic memory, answers questions across all of them with cited sources, and runs a grounded code-review swarm - entirely self-hosted on the BEAM, with every LLM call routed through an audited gateway. It also exposes recall / ask / index_repo as MCP tools, so Claude Code or any MCP client can use it from inside any repo.
| Pillar | Role in banto |
|---|---|
| gakudan | LLM backend (via sekisho) with retry/fallback; fanout review swarm with structured findings |
| bunko | pgvector store with hybrid (keyword + vector) search, DB-side metadata filtering, batched cached embedding |
| sekisho | gateway for LLM + embedding keys, budgets, cost accounting, tamper-evident audit |
| madoguchi | recall / ask / index_repo as MCP tools + repo resources, over HTTP and stdio |
| saiten | grades the review swarm in CI: self-consistency judge, JUnit, regression gate |