llanite-agent

Built for local. Inspired by the best.

An open-source agentic coding layer designed for the constraints of on-device inference. Inspired by Claude Code and OpenClaw — optimized for 7B–27B parameter models running locally.

Install llaniteView docs
llanite
Llanite · local-coder · qwen3.5:27b
> Refactor the auth module to use async/await
•
I'll start by reading the current implementation.
•Read src/auth/index.ts
└ 148 lines read.
•Edited src/auth/index.ts
└ Wrote 152 bytes to src/auth/index.ts.
•
Done. Changes made:
• Replaced all Promise.then() chains with async/await
• Added try/catch blocks for error handling
• Exported AuthError for upstream consumers
> _
esc to interrupt · /help for commandsmodel 27B/~17 GB · ◑ RAM 24.1 GB/32 GB · pressure normal

System awareness

Always knows how much headroom you have.

RAM usage, swap, and memory pressure update live in the status bar every 3 seconds. Run /ram to see a full breakdown of top processes, or /fit to check whether the loaded model fits comfortably in your available RAM.

  • Live RAM + swap in the status bar
  • Pressure indicator: normal / tight / high
  • /ram shows top processes by memory
  • /fit shows model size vs available RAM
RAM usage · top processes
21.8 GB
Ollama
348 MB
node
143 MB
CoreServices
2.9 GB
other (340)
total: 24.1 GB / 32 GB · swap: 0.2 GB
Model fit panel (/fit)
Modelqwen3.5:27b
Params27B
Disk size~17 GB
Min RAM24 GB
System RAM32 GB
Fitcomfortable

Tools & permissions

Every action is visible. Nothing runs silently.

Tool calls appear in the transcript as they happen. The permission system gives each action an explicit policy — allow it always, require confirmation, or deny it entirely.

filesystemconfirm writes

Read, write, and edit files

shellconfirm

Run shell commands

gitallow

Status, diff, log

web_fetchallow

Fetch URLs for context

Permission policy

file_writeconfirm
shellconfirm
networkdeny
secretsblock
Permission prompt (runtime)
Allow shell: git diff --staged
YesNoRevise

Capabilities

What llanite-agent brings to the table.

Live system awareness

RAM usage, swap, and pressure update every 3 seconds in the status bar. The model fit panel shows whether your model is comfortable, tight, or overloaded before and during a session.

Optimized for smaller models

Prompt strategies, context management, and planning mode are all tuned for 7B–27B parameter models. Gets more done per token than agents built for unlimited cloud context.

Explicit permission system

Every action is declared upfront — allow, confirm, or deny for file writes, shell, network, and secrets. Nothing runs without you knowing about it.

Full tool suite

Filesystem read/write, shell execution, git, and web fetch built in. All tool calls are visible in the transcript with their output previews.

Session control

Slash commands for everything.

Inspect RAM, toggle system stats, check context usage, review permissions, and control planning mode — all without leaving the prompt.

/ram
/fit
/stats
/plan on
/plan off
/context
/permissions
/help
/clear
/exit

Get started in one command.

Install a stack and llanite-agent comes with it.

$llanite install local-coder
Browse stacksView docs