Treg doubled its stars in a week
A tool gateway doubles, a video workflow climbs, and DeepSeek's new agent harness draws attention.
19-26 September 2026. Star counts come from LLMRepos daily snapshots; GitHub counts may have changed since then.
This week spans five ways to work with agents: accessing tools, making videos, coordinating coding sessions, building a harness, and running research experiments. My pick: Treg. Its shared tool gateway tackles the credential juggling that often stops agents doing useful work.
Moving fastest
superdesigndev/treg
+1,741 stars (+104%) to 3,415. Treg puts a catalogue of external services and a team's own tools behind one endpoint for agents.
Why it matters: an agent can call a tool without carrying each provider's key itself. If your team occasionally needs SEO, enrichment or other paid APIs, one shared gateway could save separate signups and keep credentials on the server.
Image: Treg README.
hypit-ai/hypit
+5,777 stars (+56.1%) to 16,083. Hypit turns a source video or written brief into an editable video workflow that an agent can run and vary.
Why it matters: the composition stays editable. If you make several versions of a short video, you can change the words, footage or effects and run it again instead of rebuilding each cut. Captions and motion graphics can render without a generation model.
Screenshot: Hypit README, showing source and rendered video side by side.
stablyai/orca
+6,366 stars (+8.8%) to 78,495. Orca coordinates coding agents in separate Git worktrees across desktop and remote runtimes.
Why it matters: one worktree per agent keeps parallel attempts isolated while their results stay in one review surface. It could help you give the same task to several agents, compare approaches, and monitor longer sessions from your phone.
Image: Orca README.
New arrival
deepseek-ai/deepseek-harness
Created 13 August; 236,238 stars by 26 September, including +6,875 in this snapshot week. DeepSeek's open-source agent harness is built around plugins.
Why it matters: its plugin-first design is worth studying if you are building an agent runtime or want to package custom behaviour. You can run its local web UI to explore the approach, but the README calls it a developer preview and warns of compatibility-breaking changes.
Quiet but still loved
karpathy/autoresearch
96,796 stars and 184 days since the last commit in the 26 September snapshot. It still gained 476 stars over the week.
Why it matters: the experiment loop is small enough to understand. An agent edits one training file, runs a fixed five-minute test, and keeps changes only when the validation score improves. It is a useful pattern for measured agent experiments; running this repo itself requires an NVIDIA GPU. The snapshot's long commit gap still matters if you plan to depend on it.
Graph: Autoresearch README. The green line tracks the best validation score so far; lower is better.
Which repo would you try first? If this reached your inbox, reply with your pick and what you would use it for.
Selection: one riser from each current-star band (500-10,000; 10,000-50,000; 50,000+), plus a repository created within 90 days and one with at least 5,000 stars and 180 days without a commit. Riser comparisons require at least 500 stars at both snapshots. Renamed or mirrored repositories with near-matching names and stars are deduplicated.
LLMRepos tracks open-source LLM and agent repositories using daily GitHub snapshots.
Get the next edition
Five open-source AI repos worth watching, every Tuesday.
Free weekly digest. Double opt-in, unsubscribe anytime.



