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

Coding agents

L80 is built to be driven by an agent: every error carries a remedy line and an exit code to branch on, --json is everywhere, and bundled skills tell the agent which command to run.

The hand-off prompt

Pick the variant that matches what you have and paste it into Claude Code, Cursor, or whatever runs your terminal. It installs L80, stops if there is no token, and comes back with the link.

Set up the L80 CLI so you can publish results to launch80, then publish a report.

1. Install it: curl -fsSL https://launch80.com/install.sh | sh
   If ~/.local/bin is not on PATH, add it, then confirm `L80 version` works.
2. Make sure uv is installed (`uv --version`); if not:
   curl -LsSf https://astral.sh/uv/install.sh | sh
3. If you are Claude Code, install the skill: L80 skills link --target claude-code
   Any other agent: run `L80 skills print` and follow that file instead.
4. Run `L80 auth status`. If it reports no token, STOP and tell me: I will save
   one to ~/.config/L80/credentials.json myself. Never ask me to paste a token
   into this chat, and never write one to a file for me.
5. Run `L80 doctor` and fix anything it flags.
6. Benchmark my server and publish the run:
   L80 betterbench --endpoint <URL> --model <MODEL> --note engine="<stack and version>" --note hardware="<GPUs>"
   Ask me for the endpoint URL, model name, engine and hardware if I have not given them.
   Use --quick only if I ask for a smoke test.
7. Give me the URL from the `published:` line and repeat every caveat and
   warning the tool printed, especially under-sampled percentiles and
   prefix-cache hits. Then run `L80 verify check <results.json>` and tell me
   whether the run could be verified by Launch80.

Agent skills

skillteaches the agent to
l80-test-reportPick a template and publish reports and BetterBench runs; check and request verification.
l80-vastRent and benchmark a GPU without leaving a box billing.
l80-pwdetScan weights, read the verdict, and publish it safely.
# copies a skill to ~/.claude/skills; Claude Code sees it on the next session start
$ L80 skills link --target claude-code$ L80 skills link --target claude-code --name l80-vast
# for any other agent: print it where that tool reads instructions
$ L80 skills print l80-pwdet
  • --name picks the skill (default l80-test-report); --force replaces an existing installation.
  • --dev --from ./skills symlinks a source checkout so edits are live.
  • L80 update refreshes every installed copy. The full text of l80-test-report is on Read before you run.

What to say

With the skill installed, the phrase picks the template:

templatesay to your agent
bench.betterbench.v1“run betterbench against my server and publish it to launch80”
bench.report.v1“publish this benchmark run to launch80”
generic.report.v1“publish these findings to launch80”
scan.pwdet.v1“scan this model with pwdet and publish it to launch80”

The skill tells the agent to repeat every warning L80 prints, to run L80 verify check before asking for review, and never to approve or reject a verification on its own judgement.

Tokens and agents