Docs
Everything you need to install, run, and customise a local AI stack with Llanite.
Install
Llanite is distributed via npm. The package exposes both llanite and the short alias lla. Node 20+ required.
$ npm install -g @llanite/cli$ llanite --version$ llanite doctorllanite doctor checks Ollama CLI, server reachability, and the registry cache. Run it first if anything looks wrong.
Quickstart
Three commands get you from zero to a running local AI agent.
$ # Install Llanite$ npm install -g @llanite/cli$ $ # Install a stack (pulls the model, configures runtime)$ llanite install standard$ $ # Launch the TUI agent session$ llanite run standardNot sure which stack fits your machine? Run llanite stacks to browse the catalog, or llanite inspect standard to see the full manifest before committing.
Local LLM Setup
Most local AI failures are integration failures, not just model failures. Llanite treats setup as a full stack: model, runtime, agentic layer, context budget, hardware fit, and policy owner.
Backend wiring
Local agents often need a provider URL, model ID, runtime process, and agent config to agree. Stacks keep those choices together so they can be inspected and cloned.
Tool calling
A model can be good at chat but poor at structured tool use. Prefer stacks that pair a tool-capable model with an agent layer known to use it well.
Context latency
Local runs slow down when the agent sends too much context. Keep context budgets intentional and clear the session when the transcript stops helping.
RAM vs VRAM
System RAM, VRAM, and Apple unified memory behave differently. Use model fit as a starting point, then watch live RAM/swap pressure during real work.
Policy ownership
Access policy belongs to the selected agent layer unless you add an outer sandbox. OpenClaw stacks are OpenClaw-managed; llanite-agent stacks are Llanite-enforced.
Repeatability
If a setup works once, it should not require rediscovery in another workspace. Clone the stack and keep the working model/runtime/agent choices together.
Policy by agent layer
llanite-agentLlanite enforces stack permissions before file writes, shell commands, network access, or denied tools run.
openclawOpenClaw manages tool visibility, sandboxing, and access through its own configuration. Llanite labels this explicitly and can package sane defaults.
custom agentEffective policy is owned by that agent unless you run it inside an outer OS/container sandbox.
Guides
Short explainers for common search questions about Llanite and local AI agent workflows.
What is Llanite?
Llanite is an open-source CLI for installing complete local AI agent stacks with models, runtimes, agents, permissions, and hardware requirements.
Local AI agents
Learn what a local AI agent is, how Llanite packages one, and why local agents are useful for private and repeatable developer workflows.
Llanite vs Ollama
Understand the difference between Llanite and Ollama: Ollama serves local models, while Llanite packages complete local AI agent stacks on top of Ollama.
Local coder: local LLM coding agents
Use Llanite local-coder to run a private local LLM coding agent with explicit models, tool access, git workflows, and permission boundaries.
Local AI agent on Mac
How to install and run a local AI agent on a Mac with Apple Silicon. Llanite and Ollama make it three commands to get a full local model stack running.
Local AI agent with Ollama
Learn how to run a local AI agent with Ollama. Llanite adds the agent loop, tools, and permissions on top of Ollama's model serving in a single install command.
Private AI coding assistant
Run a private AI coding assistant locally with Llanite and Ollama. Your code stays on your machine — no cloud API, no subscription, and three commands to get started.
Open source coding agent
Set up a fully open source coding agent with Llanite, Ollama, and open-weights models. The whole stack is inspectable, free to use, and three commands to install.
Llanite vs Continue.dev
Understand how Llanite and Continue.dev differ. Continue is an IDE extension for inline completions; Llanite is a CLI for installing and running complete local model stacks.
Commands
Full CLI reference.
Setup
llanite install <stack>Pull the model, configure the runtime, and save the stack locally.
llanite run <stack>Launch the TUI agent session for a stack.
Discovery
llanite stacksBrowse the prebuilt stack catalog.
llanite inspect <stack>Show the full resolved manifest: model, runtime, agent, tools, permissions, hardware requirements.
llanite doctor [stack]Check Llanite, Ollama, model availability, and stack compatibility.
llanite modelsList models in the registry with parameter counts and hardware requirements.
Customisation
llanite set <stack> model <model>Swap the model for a stack.
llanite set <stack> runtime <runtime>Swap the runtime for a stack.
llanite set <stack> agent <agent>Swap the agent for a stack.
llanite clone <stack> <name>Fork a stack into a custom local copy.
Cleanup
llanite remove <stack>Remove a stack config and unlink the agent.
llanite uninstall <model>Remove a model from the Ollama cache.
Registry
llanite fetchFetch and validate the Llanite registry cache.
llanite catalogList all registry entries.
llanite search <query>Search the local registry cache.
The TUI
llanite run opens a fully animated terminal UI built with Ink. It has four parts:
Status line
Pinned at the top. Shows stack ID, model name and size, runtime, agent, and workspace directory.
Transcript
Scrolling conversation history. User messages, assistant responses, and tool call results all appear here.
Working indicator
Appears while the agent is active. Shows a cycling status phrase (thinking → weighing options → deciding), a live token counter (↑ N) during response generation, and elapsed time.
System stats footer
Live RAM and swap usage, pressure level (normal / tight / high), and model size. RAM and swap have a shimmer animation to show they are refreshing every 3 seconds. Type /stats to toggle.
Slash Commands
Type any slash command in the input box during a session. Autocomplete shows matches as you type.
/helpList available slash commands.
/toolsList the built-in tools the agent can call.
/clearClear the conversation transcript.
/statusShow current session status.
/modelShow current model name, parameter count, and disk size.
/runtimeShow runtime info (Ollama URL, status).
/contextShow context window usage and remaining tokens.
/permissionsShow current permission settings for the stack.
/fitToggle the model fit panel (RAM requirements vs. system RAM).
/statsToggle the system stats footer.
/plan autoAgent decides when to plan (default).
/plan onAlways plan before every turn.
/plan offSkip planning, go straight to execution.
/ramShow RAM usage broken down by top processes, with a bar chart.
/exitEnd the session.
Agent & Planning
The Llanite agent is a ReAct-style loop: it reads your prompt, decides which tool to call (or whether to answer directly), executes the tool, observes the result, and repeats. Before each response it may optionally plan.
Planning modes
Planning produces a short step-by-step breakdown before execution. Three modes are available:
/plan autoThe agent decides whether to plan based on prompt complexity (default).
/plan onAlways produce a plan. Good for complex multi-file tasks.
/plan offSkip planning. Fastest for simple one-shot requests.
Permission prompts
When a tool requires a confirmed permission (e.g. file_write: confirm), the TUI pauses and presents a prompt before execution. You can allow once, allow for the session, or deny. Denied tools return an error the agent can observe and recover from.
Model Fit
Llanite reads your system RAM at install time and classifies each model candidate against it.
Comfortable
Model disk size is well under your available RAM. No pressure expected.
Recommended
Model fits within recommended RAM range. Good performance.
Tight
Model is at the edge of your RAM. May cause swapping under load.
Incompatible
Model requires more RAM than your system has. Not installable.
During a session, type /fit to toggle the model fit panel, which shows the full hardware breakdown for your current stack. The system stats footer shows live RAM and swap pressure.
On macOS, Ollama loads model weights into Metal GPU buffers (unified memory). These don't appear in the standard process RSS count — the /ram command accounts for this by adding the model as a separate line labelled ·Metal·.
Stack Customization
Every component of a stack can be swapped with a single command. Changes are saved immediately.
$ # Swap the model$ llanite set standard model qwen3:14b$ $ # Swap the runtime$ llanite set standard runtime ollama$ $ # Swap the agent$ llanite set standard agent llanite-agent$ $ # Fork a stack into a custom copy$ llanite clone standard my-coderAfter changing the model, run llanite doctor standard to confirm the new model is available on your Ollama instance, and llanite inspect standard to verify the updated manifest.
Concepts
Stack
The user-facing package: model, runtime, agentic layer, policy owner, context window config, and hardware requirements. Stacks are inspectable before anything runs.
Registry
A metadata layer that tells Llanite where components come from and how they fit together. Validated locally — no account required.
Runtime
The local model server. Llanite targets Ollama. The runtime is declared per-stack and can be swapped with llanite set.
Agent
The prompt loop and tool orchestration layer. Llanite ships llanite-agent and can package external agents such as OpenClaw.
Policy Owner
The layer responsible for effective tool access. Llanite-agent stacks use Llanite-enforced permissions; OpenClaw stacks use OpenClaw tool and sandbox configuration.
Doctor
A health check for Llanite, Ollama, the registry, and a specific stack's hardware compatibility. Run it before installing.
Model Fit
Llanite reads your system RAM and classifies each model as comfortable, recommended, tight, or incompatible. Shown during install and in the TUI fit panel.
Contributing
Stacks, models, runtimes, and agents are defined as YAML files in the registry. To contribute a new entry, open a pull request at github.com/lachlanforgan/llanite.
Stack schema (minimum fields)
$ id: my-stack$ name: My Stack$ summary: One-line description.$ runtime: ollama$ model: qwen3:8b$ agent: llanite-agent$ tools:$ - id: filesystem$ access: read-only$ permissions:$ file_write: deny$ shell: deny$ network: deny$ secrets: block$ hardware:$ min_ram_gb: 8$ recommended_ram_gb: 16Troubleshooting
Start with llanite doctor. It checks Ollama CLI, server reachability, registry cache, and model availability in one pass.
$ llanite doctor$ llanite doctor standard$ llanite inspect standardOllama not found
Install Ollama from ollama.com, then run ollama serve and retry llanite doctor.
Model not available
Run ollama pull <model> manually, or llanite install <stack> which pulls it automatically.
High RAM pressure during a session
Type /ram to see which processes are consuming memory. The model weights are loaded into Metal GPU buffers on macOS and appear as ·Metal· in the breakdown.
Agent not responding
Press Esc to interrupt the current turn. Type /status to check session state.
Context window full
Type /context to see usage. Use /clear to reset the conversation, or use a stack with a larger context token setting.