lmcoder

Local LLM Serving, Prompting & Sandboxed Agent Execution with lmcoder

Use lmcoder to manage local LLM serving, execute fast interactive prompts with automatic session continuity, and run sandboxed coding agents inside rootless Podman.

---

When Agents Should Use lmcoder

| Capability Level | Command | Best For | Overhead / Latency | |---|---|---|---| | L1: Fast Prompt | lmcoder prompt "<query>" | Quick Q&A, code reviews, formatting, text cleanup, unit tests | Instant (< 1s TTFT), streaming SSE | | L2: Sandboxed Agent | lmcoder agent --agent <name> | Multi-step coding tasks, full workspace refactoring, sandboxed execution | High (Container startup + agent loop) | | Model Lifecycle | lmcoder load <model> | Dynamically switching models on AMD APU / GPU VRAM+GTT | Fast (~1-3s model swap) |

---

Key Commands & Usage

1. Fast Prompting & Session Continuity (lmcoder prompt)

Send prompts directly to local models or cloud OpenAI Codex. Automatically resumes conversation history tied to the active TTY terminal session.

# Basic prompt to local active model (streaming output)
lmcoder prompt "Explain the error handling in main.go"

# Inject file context directly
lmcoder prompt -f internal/llamahost/session.go "How is the session ID computed?"
# or using inline @file syntax
lmcoder prompt "Refactor this function: @cmd/lmcoder/load.go"

# Pipe input via stdin
cat logs/server.log | lmcoder prompt "Extract error stack traces"

# Target cloud OpenAI Codex with session resumption
lmcoder prompt --host codex "Review current git diff"

# Start fresh session (ignore cached TTY history)
lmcoder prompt --fresh "Ask a clean question"

2. Sandboxed Autonomous Agent (lmcoder agent)

Run autonomous coding agents in an isolated, hardened Podman container with tool access:

# Run prime-agent against local llama-server
lmcoder agent --agent prime-agent "Implement the missing unit tests for load.go"

# Run Pi coding agent with read-only volume mounting
lmcoder agent --agent pi "Analyze codebase architecture"

Supported agents: prime-agent, pi, opencode, codex, claude.

3. Model Lifecycle (lmcoder load / unload / start)

Manage models dynamically without manually restarting daemons:

# Start server daemon (defaults to 127.0.0.1:8734 loopback)
lmcoder start

# Load or switch active model with fuzzy matching & VRAM preflight
lmcoder load qwen3-4b
lmcoder load smollm3-3b
lmcoder load 0.5b

# Unload active model to free GPU/GTT memory
lmcoder unload