Switching AI assistants shouldn't mean explaining the project again.
We shipped Relay from @LoomLabsStudio to keep the notes, decisions and next steps you choose across the assistants you connect. Not your full chat history.
https://t.co/dAM222UEta
New from Loom Labs: Relay is live.
Your assistants don't share context. ChatGPT, Claude, Cursor, Muse, Grok Bot each start from zero.
Relay keeps the projects, decisions, and next steps you choose, and makes them available across every assistant you connect. Not full chat history. Just what matters.
Try it: https://t.co/ZPDS9XTdZE
Lab notebook, not a product. The classifier only picked the tier; every ticket was still built by Claude.
Code, tests, both run reports: https://t.co/69kTfwaQXU
Experimental, clone, fork, PR.
Jev's $0.0005 of deciding cut my agents' coding bill 62%. In theory.
Same 4 tickets, 51 hidden tests, 3 ways to pick the model:
- top model on everything: $9.09
- top model picks the tier: $6.37
- classifier + fixed policy: $3.47
All three passed 51 of 51.
The surprise was the cost of deciding. Four routing calls by the top model: $1.69, half of what the cheapest setup spent on everything. The classifier (TypeSafe's Jev): $0.0005.
Limits: one run, ~15% cost noise, and a blind review still found a bug in the cheaper model's code.