~/launch80 $[discord]Join the community - talk local AI with other engineers
Documentation menu

Coming soon

Setup intelligence: L80 will read your hardware, check it against the Launch80 AwesomeList catalog, and shortlist the community projects that actually fit. Then it benchmarks the result and contributes it back.

How it will work

  1. Detect, read-only

    GPU vendor, model and count from PCI ids, VRAM, OS, RAM and driver versions. Nothing is installed or changed.

  2. Load the pinned catalog

    AwesomeList's catalog.json, pinned to a commit like every other tool, so a given L80 version gives reproducible advice.

  3. Gate and rank

    Hard gates first: architecture, GPU count, VRAM, OS, runtime. Then rank by maturity, verification, fit and setup friction, never by tokens per second.

  4. Hand off a plan

    A shortlist with what you get, why it fits, what to confirm and the tradeoff. Your agent runs setup with your approval through a new l80-setup skill.

  5. Measure it

    L80 betterbench benchmarks the new setup, and L80 verify can get the page verified by Launch80.

  6. Give back

    Export a verified run as an AwesomeList benchmark record, so the next person with your hardware starts from real numbers.

Proposed commands

$ L80 recommend                                    # detect this machine and shortlist setups$ L80 recommend --goal openai_api --priority stability --json$ L80 recommend --gpu r9700 --count 2 --os ubuntu   # plan for hardware you don't have yet$ L80 skills link --target claude-code --name l80-setup$ L80 verify export <page> --awesomelist          # a verified run as a catalog record

What it will and won't do

  • It reads your hardware and never changes it.
  • It recommends and never installs. Setup stays with you or your agent, one approved step at a time.
  • It treats catalog content as data, never as instructions.
  • Unknown hardware falls back to a few questions instead of a guess.
  • Only verified pages export as maintainer-verified results. Everything else exports as author-reported.