routstrd v0.5: Model Paths and Open Algorithms
This release introduces a new Nostr event, kind `38427`, bringing us one step closer to a Nostr native marketplace for LLM inference. If you are familiar with what makes an inference provider valuable, we invite you to be a curator.
We're one major step closer to our vision: Routing requests over Nostr.
This release introduces a new Nostr event, kind 38427. It lists the inference providers to use for a model and the exact path your request takes to reach them.
If you use vanilla OpenRouter, you're caught in a fight among its inference providers to game the routing algorithm. OpenRouter can also switch providers in the middle of a session. Each switch throws away your cached tokens, so you pay more. We're not claiming to be far better at this yet. We are taking a step toward a world where routing algorithms compete on Nostr to give users the best price. To choose whose algorithm you use, change the npub you pull the 38427 event from.
Providers *and* paths#
The current list is curated from sources including Cloudflare, Fireworks, Modal and DeepSeek. You can also choose the path to each provider. For example, you can reach Fireworks through OpenRouter, through PPQ, directly through its API, or through a Routstr node. The 38427 event's algorithm decides which path your request takes.
{"version":1,"models":{"deepseek-v4.1-flash":{"routes":[
{"url":"https://openrouter.ai/api/v1","endpoint":"fireworks"},
{"url":"https://api.fireworks.ai/inference/v1"},
{"url":"https://api.ppq.ai"}
]}}}
Here's the latest 38427 event your client uses: πnevent1qq...72yq
The current algorithm is simple. We picked paths that offer great quality at a fantastic price, including how well each one caches. A static client-side algorithm then sorts them by cost for a typical agentic request: 100k tokens in, 90% cache hit, 2k tokens out. Each path is priced on its own, so a cheap path through an expensive node can beat an expensive path through the cheapest node.
Become a curator#
We welcome more curators. If you know how LLM inference is priced, publish your own 38427 event (d tag: model-path-whitelist) and help your followers on Nostr. If a curator goes quiet, publishes a bad event or loses relay access, we need alternatives. Or even curators who compete with each other.
Right now we have model paths only for deepseek-v4.1-flash-1017. More models will come soon. It's not enabled by default yet, you can switch it on using one command:
routstrd models auto-model-path on # or off / toggle
routstrd models -m deepseek-v4.1-flash-1017 --paths # shows which path each model will take
If you see a lot of cache misses, you can ask your agent to investigate and report to
Also in v0.5#
routstrd logs: one line per request, with every retry and failover in order. Add-fto follow live and-vfor the full trace. You can also query it over HTTP:GET /requests/<uuid>.text15:57:03 alice - deepseek-v4.1-flash - 400 - 402 - 200 - 1111β¦5555- Monthly spend limits per npub: run one node for your team and cap each person's spend.
routstrd npubs add <npub> --name alice --limit-sats 50000 - Offline Cashu sends:
routstrd send 1000 --offlinebuilds a token from proofs you already hold, with no contact with the mint.--overpay-caplimits how much extra you can end up sending. - Fewer dead nodes:
providers listhides nodes that have never responded.--allshows them. - macOS fixed: if a binary was killed on launch (
Killed: 9), re-run the installer. - Plus ~50 reliability and security fixes. Payments that get stuck are now retried automatically, without a restart.
For builders: `@routstr/sdk` 0.5.1#
- Kind 38427 path lists now control
autoModelPath; it is no longer a hardcoded special case for DeepSeek. - Request tracking: each request gets a UUID and a full record of every attempt (retries and failovers) for your own app. You can store it in memory, IndexedDB, SQLite or Bun SQLite. Prompts and payment proofs are never recorded.
- Structured 402s: payment errors report exactly how many msats are required and how many are missing.
- β οΈ Breaking: when every provider fails, you now get an SDK error object (
error.type: "all_providers_failed") instead of the last node's raw response. SeeCHANGELOG.md.
Update your Client:#
routstrd update
routstr
A Decentralised LLM Protocol & Routing Marketplace Check our docs: https://docs.routstr.com Try the chat app: https://chat.routstr.com Github: https://github.com/routstr
_@routstr.com
A Decentralised LLM Protocol & Routing Marketplace Check our docs: https://docs.routstr.com Try the chat app: https://chat.routstr.com Github: https://github.com/routstr