Permissionless aggregation is an interesting phenomenon. Most new models are by default available on the Routstr network because different nodes serve models from different sources. You fetch Nostr events, you pay with sats. Models become automatically available on Routstr when any of upstreams start offering them: 1. OpenRouter 2. Venice 3. PPQ 4. Tinfoil 5. Fireworks 6. Deepseek 7. OpenAI/Anthropic of course 8. And more. (We hope to see more diversity of upstream providers) Similarly, these Abliterated models are now available from various routes. Your client -> Venice Your client -> PPQ -> Venice Happy red teaming your own tools. View quoted note →

Open weights just caught up with the closed frontier, for building and for defending! Kimi K3.1 Pro and Kimi K3.1 Pro Ultraspeed are both live on the Routstr network. Building: #4 in Artificial Analysis Intelligence Index. #2 in Terminal Bench 4.0, only behind Opus 5.5. Cyber Defense: #2 in AA Cyber Index, but #1 among publicly available models. And it doesn't refuse the work. Try them, model: kimi-k3.1-pro / kimi-k3.1-pro-ultraspeed Happy building and defending.

LiteLLM was compromised again, your Routstr node is not affected. A malicious build of litellm 1.107.0 shipped last week with a credential stealer and remote-code loader in all seven wheels. It's the second litellm hijack this year. Routstr is unaffected. We pin litellm >=1.101.2,<1.102 with an exact lockfile, so 1.107.0 can't resolve in a Routstr deployment, and we found none of the published indicators in our environments. If you are running your own venv as a developer, check "pip show litellm". If it says 1.107.0, remove cook.pth, cook.py, and sitecustomize.py from site-packages before running pip again, startup re-triggers the payload, then treat every secret on that host as compromised and rebuild. We're removing litellm. We only use it as a translation layer, and two hijacks in one year is not a good sign.

X (formerly Twitter)

Aikido (@AikidoSecurity) on X

❗️LiteLLM was compromised (Again). The latest version 1.107.0 contains an infostealer and RAT. All 7 compiled wheels are backdoored. Anyone wh...

Periodic reminder to update your client to get the best prices and cache hit rates. View quoted note →

New Release: Routstr Core 0.5.0 A Major New Feature: Modular Model Path Control. With this new release, any client can specify the exact upstream inference provider they want to use for a particular model. Including all the inference providers on OpenRouter. Second Major Feature: Upstream Provider Test Suite You can test if everything is good with an upstream inference provider you just added. Other changes in: 1. Intelligence: • /v1/systemone — typed AI decisions, paid per call (Jev) • /v1/decisions — OpenAI's new Decisions API, proxied • native Venice + DeepSeek 2. Money: • Lightning refunds + payouts in tx history • crash-safe refund claims 3. Ops: • one-command private node onboarding and llms.txt for agent powered onboarding • faster requests, provider cooldowns, never strand a reservation ~100 PRs. 254 files. 4 weeks.

GitHub

Release v0.5.0 · Routstr/routstr-core

What's Changed fix(ehbp): pass through enclave key-config 422 as problem+json by @sh1ftred in #714 fix(ehbp): raise upstream timeout to 60s and re...

Not just a problem with secondary markets for AI inference, it solves this problem for anything that is resold online using an API key. Please help us battle test it! View quoted note →

Unreal times, an idea from a conversation here on Nostr to full implementation in 2-3 days. Juraj is a juggernaut! Open ideas, open execution. Open source will be unstoppable. Grateful for this community. ❤ View quoted note →

The upcoming Private Provider App goes head-to-head with Codex Desktop and Claude CoWork but it's built entirely on Nostr rails, so it'll be 10x better, with none of the legacy headaches like API keys, GitHub connections, Google 🤧, etc. But unlike Codex/Claude, you aren't leaking personal information because of TEEs and you only need to topup using some Bitcoin! Plus, everything is on your phone so you can build while playing with your kids and touching grass. 🧡 Super excited! View quoted note →

Jev is live on Routstr! Jev jev jev jev jev, I'm sure you've heard of this by now, it's a new model architecture that is super fast and super cheap and works as classifier for anything you need. Look here for more inspiration:

GitHub

GitHub - cobanov/awesome-jev: A curated, source-backed list of projects built with Jev, TypeSafe AI's System One model for typed decisions.

A curated, source-backed list of projects built with Jev, TypeSafe AI's System One model for typed decisions. - cobanov/awesome-jev

You can finally jevelop (sorry 😂) using jev using nothing but a couple of sats! Update your routstrd if you want to use it through it. Or you can use it through our sdk as well. (pnpm install @routstr/sdk) The model is "jev-latest" live on one node rn:

Routstr

Routstr model management

Give this skill to your agent and start building using this, just tell it to use and query

TypeSafe

Agent skill - TypeSafe

Drop-in skill for Claude Code, Codex, and other agent environments.

We Sunset Our Official Node. Zero Downtime. All nodes are community-run now 🧡 It's been a month and a half since we sunset api.routstr.com, the last "official" Routstr node. Nobody noticed a thing. routstrd and routstr-sdk discover nodes on Nostr and pick the cheapest one for you automatically. No one can rugpull you. No one can capture you. Your client truly works for you. We're removing the team as a dependency. Point out any centralization left and we'll remove it next 🔨