Skip to content
ignisignis

Install with a coding agent

A coding agent that can run commands and read web pages (opencode, Claude Code, Codex and the like) can install ignis for you. The repository’s README points an agent, on its first screen, to a page written for it: the agent path, with the checks to run, the choices that are yours and the commands of each step.

What to tell it

Section titled “What to tell it”
Install ignis from https://github.com/gpillon/ignis

The same page answers the other things you might ask it:

Test a 16 GB card

Section titled “Test a 16 GB card”

The rtx50-16gb profile has never run on a real 16 GB card. With an RTX 5060 Ti 16 GB, a 5070 Ti or a 5080, your agent can install ignis, test it and write a report for issue #360:

I have a 16 GB RTX 50 card. Read https://raw.githubusercontent.com/gpillon/ignis/main/docs/user/paths/agent.md and follow its section on testing a 16 GB card: help me install ignis, test it, and write the report. Ask me before downloading or installing anything.

It shows you its plan first, asks before it downloads the 13.4 GB model, tries the default and then a few settings if the default is refused, and writes the report in English for you to read before anything is posted. An advanced part, about 30 minutes, measures the context ceiling, one lane against two and DFlash2 on against off, if you want it. At the end it asks whether to keep ignis or remove what it installed (the agent path’s test).

What it does

Section titled “What it does”
  1. Checks the machine: the card, the driver, the system, the disk and, for Flash-Next, the host RAM. On a card ignis does not support it stops and tells you why (What it runs on).
  2. Asks you which model, where to keep it (tens of GiB), the release archive or the container, and whether the server should be reachable from other machines.
  3. Installs: downloads the latest release, checks its SHA-256, fetches the model you chose into the directory you chose, starts the server and waits until /ready answers.
  4. Connects a client when you ask, itself included: ignis is an OpenAI-compatible endpoint (Connect a client).

What it leaves to you

Section titled “What it leaves to you”

The agent path tells it to ask before any of these:

  • downloading a model;
  • installing a driver or the CUDA runtime, or anything that needs administrator rights;
  • stopping a process that holds the card;
  • making the server reachable beyond this machine;
  • building from source instead of taking a release;
  • posting a report, or anything else, to GitHub;
  • writing a config file, or deleting what it installed.

It matters for the download in particular: started from an agent’s shell, with no terminal to answer, the server fetches its model without asking (Letting the server fetch it).