localcoder 0.1.0

Local AI coding agent — auto-installs, auto-serves, zero config. Works with Gemma 4, Qwen 3.5, and any model via llama.cpp or Ollama.

localcoder

The local coding CLI that does the obvious things nobody else does.

pipx install localcoder

I wanted to paste a screenshot into my coding assistant and see it inline. No tool did that locally. So I built one.

Cost: $1.30/month vs $110/month

Running local saves 85-141x compared to cloud APIs:

Usage	Claude Sonnet	Claude Opus	Local (US)	Local (India)
4h/day	$55/mo	$91/mo	$0.65/mo	$0.29/mo
8h/day	$110/mo	$183/mo	$1.30/mo	$0.58/mo
10h/day	$137/mo	$228/mo	$1.62/mo	$0.72/mo

Based on: Gemma 4 26B at 47 tok/s, 30% active generation, M4 Pro 30W. Electricity: worldpopulationreview.com. API: anthropic.com.

Annual savings: ~$1,300-$2,700 depending on usage and API choice.

What's Actually Different

Feature	localcoder	aider	OpenCode	Claude Code
Paste image, see it inline	Ctrl+V → shows in terminal	no	no	cloud only
Voice input (local)	Ctrl+R → Whisper, no cloud	no	no	no
See GPU memory while coding	/gpu → live stats	no	no	no
Computer use (screenshot + click)	built-in	no	no	cloud only
Free GPU when it's slow	/clean → before/after	no	no	n/a
Browse HuggingFace models	built-in model browser	no	no	n/a
Works offline	100%	partial	partial	no
Cost	$0.00	API costs	API costs	$20/mo+

Demo

❯ localcoder

  localcoder  ·  local AI coding agent  ·  $0.00 forever

  ┌──────────────────────────────────────────────────┐
  │  LOCAL CODER                                     │
  └──────────────────────────────────────────────────┘

  ● Gemma 4 26B Q3_K_XL  ·  llama.cpp  ·  128K  ·  ● GPU  ·  47 tok/s
  ✓ offline  ·  no API keys  ·  no data sent

  ctrl+r voice  ctrl+v image  /gpu stats  /clean free  /models switch

❯ /gpu
  GPU  ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━  12/16GB  3GB free
  Swap 3GB  Pressure normal
  Model Gemma 4 26B Q3_K_XL  GPU  ctx 128K  footprint 2311MB

Benchmark — M4 Pro 24GB

Real tests, real hardware, no synthetic benchmarks:

Model	Size	tok/s	Notes
Gemma 4 26B Q3_K_XL	12.0GB	47	Best overall — vision + tool calling
Qwen3.5-35B MoE Q2_K_XL	11.3GB	46	Best coding quality
Qwen3.5-4B Q4_K_XL	2.7GB	46	Quick tasks
Gemma 4 E4B Q4_K_M	5.0GB	56	Fastest — good for 16GB Macs
Qwen3.5-27B Dense	13.4GB	7	Swap thrashing — don't use on 24GB

Install

# macOS (Apple Silicon)
pipx install localcoder

# First run — auto-detects hardware, shows what fits, starts model
localcoder

Needs llama.cpp or Ollama. First run wizard handles this.

Commands

localcoder                           # interactive coding
localcoder -p "build a react app"    # one-shot
localcoder --yolo                    # auto-approve tools

While Coding

Command	What
Ctrl+V	Paste + display image from clipboard
Ctrl+R	Toggle voice input (local Whisper)
/gpu	GPU memory, swap, model status
/clean	Free GPU memory with before/after
/models	Switch model (includes HuggingFace trending)
/clear	Clear conversation

Also works with Claude Code

Don't want localcoder's agent? Use Claude Code with your local model instead:

pip install localfit
localfit --launch claude --model gemma4-26b

One command: starts model → configures Claude Code → launches with --bare flag. See localfit for details.

GPU Toolkit (localfit inside)

localcoder --simulate               # will this model fit my GPU?
localcoder --fetch unsloth/...      # check all quants from HuggingFace
localcoder --bench                  # benchmark models on YOUR hardware
localcoder --health                 # GPU health dashboard
localcoder --config opencode        # auto-configure OpenCode for local models
localcoder --config aider           # auto-configure aider

Also available standalone: pipx install localfit

Hardware

Mac	RAM	Best Model	Speed
Air M2	8 GB	Qwen 3.5 4B	50 tok/s
Air M3	16 GB	Gemma 4 E4B	57 tok/s
Pro M4	24 GB	Gemma 4 26B Q3_K_XL	47 tok/s

License

Apache-2.0

Security

Sandbox mode is ON by default. Protects against destructive model outputs:

Blocked	Examples
Destructive commands	rm -rf, sudo, kill, mkfs
Pipe to shell	curl ... | bash, wget ... | sh
Protected paths	~/.ssh, ~/.aws, ~/.env, /etc/
Path traversal	../../etc/passwd
Computer use	Disabled in sandbox

localcoder                    # sandboxed (default)
localcoder --yolo             # auto-approve but sandbox ON
localcoder --unrestricted     # sandbox OFF (shows warning)

Approved tools are remembered across sessions (~/.localcoder/approved_tools.json).

Tests

pip install pytest
pytest tests/ -v      # 19 tests