QuorumRouter is a source-available Deno framework with two distinct multi-model paths. Best Route fans prompts out to independent model adapters, validates every response with Zod, requires quorum, and synthesizes the strongest answer. Agent Chat gives different models shared conversation context so they can challenge, revise, and converge across visible turns. Any repository mutation still requires external SafeLoop authority, exact-digest approval, watched execution, and verified artifacts.

Hey Product Hunt 👋 I built QuorumRouter because multi-model AI systems often blur two fundamentally different ideas: 1. asking several models independently and selecting the best answer; 2. letting different models actually talk, disagree, revise, and converge. QuorumRouter keeps those paths separate. Best Route runs model candidates in isolated contexts, validates every response, requires quorum, compares the results, and synthesizes the strongest final answer. Agent Chat invokes two distinct working provider/model wrappers and passes bounded conversation context between them. The CLI displays every model identity, reply target, and response as it arrives—so you can see one model challenge another, cause a revision, and move toward a conclusion. If fewer than two working model identities can be established, Agent Chat fails closed rather than pretending a single-model response is a conversation. Here is the real CLI path: npx --yes github:sakamoto-sann/quorum-router#main my-quorum-router cd my-quorum-router deno task intake deno task models:list RUN_EXTERNAL_MODEL_DOGFOOD=1 \ RUN_EXPERIMENTAL_AGENT_CHAT=1 \ QUORUM_ROUTER_AGENT_CHAT_MAX_TURNS=6 \ deno task agent-chat \ --prompt "Debate the safest migration plan" I verified the live path with a four-turn conversation between: OpenAI/codex-cli ↔️ Cognition/devin-cli Each turn preserved model identity and replying to lineage, passed schema validation and redaction checks, and was recorded in out/agent-chat-trace.json. When a conversation produces an action with side effects, the models still do not receive execution authority. The exact proposal goes to SafeLoop, which handles policy checks, digest-bound approval, watched execution, artifact verification, and audit receipts. If approval or verification fails, the workflow stops. The launch videos are deterministic visualizations rather than recordings of live model API traffic. The live wrapper-backed Agent Chat command is implemented separately and available from the source-backed scaffold. MIT-licensed open source. Commercial and production use are permitted. I’d especially appreciate feedback on: - whether Best Route and Agent Chat feel clearly differentiated; - which provider combinations you want next; - what kinds of agent actions should always fail closed; - how you would use visible cross-model debate in your own workflow. Thanks for checking it out—I’ll be here answering questions throughout the launch! 🚀

Spent an afternoon wiring QuorumRouter into a side project and the Zod validation step before quorum approval actually caught a flaky adapter I would have shipped. Curious how the SafeLoop approval flow feels once you stack more than a few repos.

Spent a couple hours wiring QuorumRouter into a small Deno script and the Zod validation step before quorum really does catch the lazy model. The SafeLoop approval flow feels heavy at first, but seeing the exact-digest check actually block a bad edit was a nice moment.

QuorumRouter

Fail-closed routing and safe execution for AI agents

Fail-closed routing and safe execution for AI agents