🚀 Quick start¶

Goal: installed binary → “oh, it edited my file” in about a minute. ⏱️ Secondary goal: smile at least once. 😄

1. 🔌 Point at a model === "🍎 oMLX"

   omlx start
   slmcode doctor
   

   Green doctor = green light. Red doctor = “have you tried turning the model on?” === "🦙 Ollama"

   ollama serve
   slmcode config set provider ollama
   slmcode config set model qwen2.5-coder:14b
   slmcode config set endpoint http://127.0.0.1:11434
   slmcode doctor
   

   === "🌌 Other" Flags, env, Studio Settings, or a stack:

   slmcode stack list
   slmcode stack apply deepseek   # needs DEEPSEEK_API_KEY or /auth set
   slmcode doctor
   

   Details → Providers. Bring your own brain.
2. 🛝 Init a playground

   mkdir -p /tmp/slm-demo && cd /tmp/slm-demo
   printf 'package main\n\nfunc Hello() string { return "hi" }\n' > hello.go
   printf '# Agents\n\nPrefer tiny Go edits and clear godoc comments.\n' > AGENTS.md
   slmcode init              # detects Go and applies the go pack for you
   

   init prints the pack it picked (pack: go (detected)). Override it with slmcode blocks apply <pack> if you disagree — slmcode blocks list shows all thirteen. Edit .slmcode/PROJECT.md with two honest sentences about the stack. (Lying to your own memory file is a bold strategy. We don’t recommend it.)
3. 🎯 Run the pipeline

   slmcode run -v "Add a Go doc comment to Hello() explaining it returns a greeting. Keep it tiny."
   

   Pass checklist ✅ - [ ] the run ends with a Changes block naming hello.go and a +N −M count - [ ] hello.go has a real // Hello … comment - [ ] slmcode board shows completed work - [ ] slmcode session list has a run - [ ] .slmcode/SKILLS.md exists (the flywheel sneezed) Got ⚠ no files changed instead? That is the harness telling you the truth: the model claimed an edit it never made, and the evidence gate refused it. The Next lines under it are the ones to follow — slmcode task show T1 prints the reviewer's verdict, the gate that refused the task and the (empty) diff of its focus files. Smaller scope and a sharper acceptance line fix this more often than a bigger model does.
4. 🕹️ Open a cockpit

   slmcode studio    # clicky mode — open the URL it PRINTS
   # and/or
   slmcode           # premium TUI — keyboard mode
   

   Pick your fighter. Both talk to the same harness. 🥊 !!! warning "Open the URL Studio prints, not http://127.0.0.1:7420" Studio can read this repo, rewrite its config, store your API keys and start runs, so it is behind a per-launch session token: the URL it prints looks like http://127.0.0.1:7420/?t=8f3c…. A bare http://127.0.0.1:7420/ gets a 401 page telling you to go back to the terminal. Opening the tokenised URL once mints a cookie and the token stops appearing in the address bar. → Studio security model !!! note "Built from source? Run make bootstrap first" A released binary ships the Studio SPA. A binary you built yourself with go build does not — the SPA is a Vite app in web/ that has to be built into cmd/slmcode/ui/ before go:embed can pick it up. Without it, slmcode studio starts normally and serves a page saying the UI has not been built, and prints the same on startup. Fix:

   make bootstrap      # installs web/ npm deps (needs Node 22+), then builds the UI
   make build
   

   web/package-lock.json is currently out of date with web/package.json, so npm ci cannot run; make bootstrap says so and falls back to npm install, which regenerates the lock — commit it. → Studio: building the UI · Troubleshooting

Level-ups 🎮¶

# Safer on real repos (stage first, apply later)
slmcode config set permission review
slmcode run -v "Add a unit test for Hello()"
slmcode apply       # interactive: a/s/e/v/r/A/q per file (--all to apply everything)

# Help small models think twice (literally)
slmcode config set think_passes 2
slmcode config set retries 2

# One specialist, no full circus
slmcode run --agent explorer "Where is Hello defined?"

Where next? 🗺️¶

☀️ Made with ♥ by UnicoLab