← ipsupport-code
how to use it

✦ zero to running in two minutes

Install it, point it at a model — local or cloud — and get it writing and running real code in your terminal. This page covers the basics; the README has the full reference for everything else (sub-agents, MCP, skills, sessions…).

1 install

One line. No Python, no Node, no runtime to manage.

macOS / Linux
$ curl -fsSL https://raw.githubusercontent.com/ipsupport-llc/ipsupport-code/main/scripts/install.sh | sh
Windows (PowerShell)
PS> iex (irm https://ipsupport-llc.github.io/ipsupport-code/install.ps1)

Verifies the SHA-256 and drops one binary in ~/.local/bin. Prefer a manual download? Grab an archive from releases.

2 connect it to a model

Pick one. You can switch any time with /ai — the session, tokens, and mode all carry over.

Local — LM Studio (the default)

Open LM Studio, download a model, load it, and start its local server (the "Developer" tab → Start Server). That's it — ipsupport-code talks to localhost:1234 out of the box, nothing to configure.

pick which loaded model to use
❯ /model              # lists what's loaded in LM Studio
❯ /model nemotron     # switches to it (a substring is enough)
recommended

For the best tool-calling reliability, use a model fine-tuned specifically for this agent: Nemotron-3.5-Lightning-30B-A3B — ipsupport-code LoRA. Trained on ipsupport-code's own tool-call schema, so it emits the exact {"action","params"} shape reliably instead of garbling it. Any instruct model in LM Studio's catalog works, though — this one is just tuned for less friction.

Local — Ollama (or any keyless OpenAI-compatible server)

No command needed — add it straight to the config, no API key required:

~/.config/ipsupport-code/config.json
{ "provider": "ollama",
  "providers": { "ollama": {
    "base_url": "http://localhost:11434/v1",
    "model": "nemotron-3.5"
  } } }

Then /ai ollama to switch to it (or /config to do the same thing from the settings panel).

Cloud — OpenAI, Anthropic, Grok, Groq, OpenRouter, Z.ai

One command adds a key and you're switched:

inside ipsupport-code
❯ /ai key openrouter sk-…
❯ /ai openrouter
❯ /model sonnet    # filters the catalog — great for OpenRouter's 300+

Keys are written chmod 600 to ~/.config/ipsupport-code/config.json, or fall back to the usual env var (OPENAI_API_KEY, ANTHROPIC_API_KEY, …).

Don't want to remember the exact syntax? /config is the same thing as a menu — ↑↓ to a field, Enter to cycle or jump into it (provider, model, key, permissions, color…), esc to close. It saves as you go, no JSON to hand-edit.

3 hello world, on a fresh MacBook

The whole loop, start to finish.

  1. Install: run the curl one-liner above in Terminal.
  2. Start a model: open LM Studio, load any instruct model (the recommended one above, or whatever's already downloaded), hit Start Server on the Developer tab.
  3. Run it: mkdir ~/hello && cd ~/hello && ipsupport-code
  4. Ask it something: type write hello world in Go and run it and press Enter.
~/hello
❯ write hello world in Go and run it
  ⚙ file write  main.go
  ⚙ run   go run main.go
  → exit 0   Hello, world!
  Wrote main.go and ran it — printed "Hello, world!".
  NEXT: add a go.mod

Want the same task on a cloud model instead? Add a key and switch first — the task itself doesn't change:

~/hello
❯ /ai key openai sk-…
❯ /ai openai
❯ write hello world in Go and run it
  ⚙ file write  main.go
  ⚙ run   go run main.go
  → exit 0   Hello, world!

✈ works with zero signal

On a plane, out of coverage, on a train through a tunnel — none of that matters when the model lives on your own machine. LM Studio, the model, and the agent are all local; there's no cloud round-trip to lose.

no wifi, no problem
❯ /offline on   # skip update checks, refuse the web tool cleanly
❯ fix the failing test in payments.go
  ⚙ file read  payments.go
  ⚙ run   go test ./...
  → exit 0   ok

/offline only stops the agent from reaching out (the web tool, update checks) — it doesn't touch your model connection. Pair it with a local LM Studio setup and the whole loop — read, edit, run, test — needs no connectivity at all.

5 the essentials

A handful of commands cover most of what you'll actually reach for day to day. Type /help any time for the full list.

/goal <text>

For anything bigger than one turn. A judge checks whether it's actually done before accepting the finish, and re-feeds it if not.

/goal add tests for the API and make them pass

/cd <dir>

Change the working directory the agent operates in — everything (files, commands, git) stays inside it, a hard jail by default.

/cd ../other-project

/reasoning <level>

Dial a reasoning-capable model's thinking effort — helps a slow local model answer faster, or a tricky task think harder.

/reasoning high

/model

List or switch models on the current provider — filters as you type, handy on a big catalog like OpenRouter.

/model nemotron

/plan · /auto

Plan mode proposes without touching anything; auto executes. shift+tab flips between them mid-conversation.

shift+tab

/status

One screen: provider, model, workspace, jail, permissions, session, and the standing goal if you have one.

/status

/config

All of the above (provider, model, key, permissions, color…) as an arrow-key menu instead of commands to memorize. Saves as you go.

/config

→ go deeper

Sub-agents (delegate to another model), MCP servers, skills, budgets, sessions, rewind, and the full command reference are all in the README.