Hardware Entrepreneur. Wannabe Cryptographer. Former EIR at Matter Labs. Jevlinist @JevlinAI
San Francisco, CA
Joined March 2019
- Tweets146
- Following80
- Followers208
- Likes46
Pinned Tweet
This is what a CLI for continual learning looks like.
rlcli trains your agent’s model with fused GSPO on your own GPUs — one forward_backward("gspo") call, 23% faster steps.
Traces in, weights out. Open source.
Most SF Startups are trying to crack continual learning from the harness…
But that’s only because they still think RL is either too hard or too expensive .
Bring traces from LangSmith, chat logs, or your own environments.
Runs on your local Docker — no cloud sandbox account.
Nothing leaves your machine.
No more duct-taping a training API to an RL engine.
Repo: github.com/polygramme/rlcli
pip install polygramme-rlcli
Gaurav Shukla retweeted
Long-lived API credentials shouldn’t live inside coding agents or sandboxes.
In this demo, an agent tries to post to Linear, gets blocked, then uses OpenCloak to perform an RFC 8693 token exchange, get a short-lived scoped token, and complete the action.
My co-founder @RunnerElektrode built an open-source prototype. Repo in reply.
Introducing OpenCloak: Auth0-style OAuth Vault for @openclaw
Keeps Discord/Github/Slack credentials outside of OpenClaw. Uses @Tailscale's tsidp for agent identity and implements a RFC 8693 token exchange.
Self hosted alternative to @auth0 vault
The following could be extended to self-hosted MCP auth solutions, acting as a token exchange server for scope-based session management.
Support for @okta @WorkOS @descopeinc is a no-brainer.
code is open source here: github.com/runnerelectrode/o…
If you're deploying locally, you'll still need Tailscale.
I was experimenting with deploying multi-instance openclaw on a single gateway over the weekend and used Claude Opus 4.6 to deploy it on DO and railway and the experience was really time consuming.
So I built Spawn to deploy and self-heal any Github repo ↓
