ButterflyOS forks your production graph, takes down any combination of services and people, and lets a swarm of agents compute the blast radius, find the knowledge that just walked out the door, and heal the organization, all in milliseconds.
Why graphs
“What breaks if this service goes down?” and “What happens if our lead architect quits?” are traversals, not text searches. Flat retrieval can't compose the edges that matter.
Payments → Ledger → Token KV is not in any document. It exists only as the composition of two dependency edges. PageRank names the chokepoint; multi-hop Cypher gives the exact failure path.
The one warning about shard 14 sits in a Slack thread linked to a commit, a repository and a service. Components analysis finds the orphaned subgraphs; agents rebuild the context and pick a successor.
How it works
GRAPH.COPY clones production into an isolated sandbox tenant in tens of milliseconds. Production is never touched.
Multi-hop Cypher and algo.pageRank trace the cascade. algo.betweenness and algo.WCC expose ownerless subgraphs.
Four agents call graph tools over MCP, reconstruct tribal knowledge, and rewire ownership with auditable edges.
The episode is distilled into a procedural playbook. The next identical failure is recalled with a single query.
The swarm
No conclusion is accepted unless every node in its provenance path exists in FalkorDB. No Node, No Action.
Maps the failure cascade, ranks chokepoints, totals RPS at risk and breached SLAs.
Finds systems whose only expert is gone and flags code that contradicts policy.
Walks commits, Slack and ADRs to write a cited handover for every orphan.
Scores successors in Cypher and rewires ownership, blocked by policy where needed.
Built for all three tracks
Seven MCP tools. Every action is chosen by graph topology and returns its Cypher, latency and provenance.
Episodic hand-offs, live semantic topology and procedural playbooks, in isolated multigraph tenants.
People, repos, commits, decisions and chats in one graph. Ask who owns what and trace every answer to its source.
Architecture
The graph decides what is true. A pluggable language model only phrases it, and is rejected the moment it cites a node that doesn't exist.
Pick a scenario, press simulate, and watch the organization heal itself.