Back to docs

Llanite guide

Local coder: local LLM coding agents

A local coder is a local LLM coding agent that uses a model served on your machine alongside tools for file access, shell execution, and git. Unlike a cloud coding assistant, a local coding agent keeps all code and context on your machine and can work without an internet connection once the model is downloaded.

What a local coding agent is

A local coding agent combines a locally served model with tools that allow it to read and write files, run shell commands, inspect git history, and navigate a codebase. The agent uses these tools in a loop — reading context, taking actions, observing results — until a task is complete. The key advantages over cloud coding assistants are privacy (code never leaves your machine), cost (no per-token API fees for long sessions), and offline capability (works without an internet connection).

What a coding stack contains

A Llanite coding stack declares every component of the local agent setup: the model (usually a code-optimized model such as Qwen 2.5 Coder), the local runtime (Ollama), the agent layer (llanite-agent or OpenClaw), the tools available to the agent (filesystem, shell, git), the permission policy for each tool, and the minimum and recommended RAM for the setup. Having all of this in one inspectable manifest means you can review what the agent is allowed to do before it runs.

Available coding stacks

Llanite ships four pre-configured stacks. The Standard stack (local-coder) uses a 27B Qwen 2.5 Coder model with OpenClaw for coding tasks on machines with 32 GB or more. The Compact stack (repo-reviewer) uses an 8B model in read-only mode for code review on 16 GB machines. The Lite stack (lightweight-coder) uses a 9B Qwen model for coding on 16 GB machines with write access. The Chat stack (private-chat) uses Gemma 3 with no tools for private conversation. Run llanite stacks to see all available options with RAM requirements.

Getting started

Setting up a local coding agent requires Ollama and the Llanite CLI. Install both, run llanite doctor to confirm they are working, then install the coding stack that fits your machine's RAM.

$ llanite install local-coder$ llanite run local-coder
  1. 1Install Ollama from ollama.com and start it: ollama serve
  2. 2Install Llanite: npm install -g @llanite/cli
  3. 3Check prerequisites: llanite doctor
  4. 4Browse stacks and pick one for your RAM: llanite stacks
  5. 5Install the stack: llanite install local-coder
  6. 6Start the agent session: llanite run local-coder

Common workflows

Local coding agents handle a range of development tasks. Code review and explanation work well even on smaller models — the agent can read files and describe what they do without needing to write anything. Refactoring, debugging, and test generation benefit from larger models with write access. Repository exploration is a good starting task because it requires only read access and produces useful output quickly. Documentation updates and README generation are low-risk starting points for testing write permissions on a new stack.

Safety and permissions

Before a local coding agent can write files or run shell commands, those permissions must be explicitly enabled in the stack manifest. Llanite displays the full permission set when you run llanite inspect before installing, and shows active permissions in the TUI during a session with /permissions. Tool calls that require confirmed access will pause and ask before executing. This means you can see and control exactly what the agent is permitted to do without relying on the model to self-limit.

FAQ

What is the best local coding agent stack?

The best stack depends on your machine's RAM and your task. The Standard stack (local-coder) with 27B Qwen 2.5 Coder produces the strongest coding output but needs 32 GB or more. The Lite stack (lightweight-coder) with 9B Qwen works well on 16 GB machines for most coding tasks.

Can local coding agents work offline?

Yes, once the model is downloaded. The initial llanite install pulls the model via Ollama, which requires an internet connection. After that, agent sessions run fully locally. Some optional tools like web search still require network access.

What tools does a coding agent use?

Coding stacks typically enable filesystem read and write (for reading and editing files), shell execution (for running tests and build commands), and git (for reading history and staged changes). The exact tool set is declared in the stack manifest and visible with llanite inspect.

How do I see what permissions my stack has?

Run llanite inspect <stack> before installing to see the full permission manifest. During a session, type /permissions in the TUI to see the current active permissions for the running stack.

Can a local coding agent edit files?

Yes, if the stack grants filesystem write access. Read-only stacks like repo-reviewer cannot modify files, which makes them safe for reviewing unfamiliar codebases. Write-capable stacks like local-coder and lightweight-coder will confirm before writing if the permission is set to confirm rather than allow.

What models work best for coding tasks?

Code-optimized models like Qwen 2.5 Coder 7B, 14B, and 32B perform well for coding agent tasks. They produce better tool calls and more accurate code edits than general-purpose models of the same size. Llanite's standard coding stacks pair these models with agents and tool configurations tuned for coding workflows.