localcoder 0.1.0__py3-none-any.whl

localcoder/voice.py

@@ -0,0 +1,187 @@
+"""Voice input — hold Space to talk, release to transcribe."""
+import os, subprocess, signal, tempfile, time, shutil
+from pathlib import Path
+
+from rich.console import Console
+
+console = Console()
+
+# ── Paths ──
+WHISPER_BIN = shutil.which("whisper-cli") or "/opt/homebrew/bin/whisper-cli"
+WHISPER_MODEL_DIR = Path.home() / ".local/share/whisper"
+WHISPER_MODEL = WHISPER_MODEL_DIR / "ggml-base.bin"
+SOX_REC = shutil.which("rec") or "/opt/homebrew/bin/rec"
+
+
+def is_voice_available():
+    """Check if voice input dependencies are installed."""
+    return os.path.exists(WHISPER_BIN) and WHISPER_MODEL.exists() and os.path.exists(SOX_REC)
+
+
+def check_mic_permission():
+    """Check macOS microphone permission by attempting a quick recording."""
+    try:
+        tmp = tempfile.mktemp(suffix=".wav")
+        proc = subprocess.run(
+            [SOX_REC, "-q", "-r", "16000", "-c", "1", "-b", "16", tmp, "trim", "0", "0.5"],
+            capture_output=True, text=True, timeout=5
+        )
+        if os.path.exists(tmp):
+            size = os.path.getsize(tmp)
+            os.unlink(tmp)
+            if size > 100:
+                return True
+        # Check stderr for permission errors
+        if "permission" in proc.stderr.lower() or "not authorized" in proc.stderr.lower():
+            return False
+        return proc.returncode == 0
+    except:
+        return False
+
+
+def setup_voice():
+    """Interactive setup for voice input."""
+    console.print(f"\n  [bold magenta]Voice Input Setup[/]\n")
+
+    # Step 1: Check sox/rec
+    if not os.path.exists(SOX_REC):
+        console.print(f"  [yellow]Installing sox (audio recorder)...[/]")
+        r = subprocess.run(["brew", "install", "sox"], timeout=120)
+        if r.returncode != 0:
+            console.print(f"  [red]Failed to install sox. Run: brew install sox[/]")
+            return False
+        console.print(f"  [green]✓ sox installed[/]")
+    else:
+        console.print(f"  [green]✓ sox already installed[/]")
+
+    # Step 2: Check whisper-cli
+    whisper_bin = shutil.which("whisper-cli")
+    if not whisper_bin:
+        console.print(f"  [yellow]Installing whisper-cpp...[/]")
+        r = subprocess.run(["brew", "install", "whisper-cpp"], timeout=300)
+        if r.returncode != 0:
+            console.print(f"  [red]Failed to install whisper-cpp. Run: brew install whisper-cpp[/]")
+            return False
+        console.print(f"  [green]✓ whisper-cpp installed[/]")
+    else:
+        console.print(f"  [green]✓ whisper-cpp already installed[/]")
+
+    # Step 3: Download whisper model
+    if not WHISPER_MODEL.exists():
+        console.print(f"  [yellow]Downloading whisper base model (148MB)...[/]")
+        console.print(f"  [dim]Supports: English, French, Arabic + 95 more languages[/]")
+        WHISPER_MODEL_DIR.mkdir(parents=True, exist_ok=True)
+        r = subprocess.run([
+            "curl", "-L", "--progress-bar",
+            "-o", str(WHISPER_MODEL),
+            "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-base.bin"
+        ], timeout=300)
+        if r.returncode != 0 or not WHISPER_MODEL.exists():
+            console.print(f"  [red]Failed to download whisper model[/]")
+            return False
+        console.print(f"  [green]✓ Whisper base model downloaded ({WHISPER_MODEL.stat().st_size // 1024 // 1024}MB)[/]")
+    else:
+        console.print(f"  [green]✓ Whisper model ready ({WHISPER_MODEL.stat().st_size // 1024 // 1024}MB)[/]")
+
+    # Step 4: Check microphone permission
+    console.print(f"\n  [dim]Testing microphone access...[/]")
+    if check_mic_permission():
+        console.print(f"  [green]✓ Microphone access granted[/]")
+    else:
+        console.print(f"  [yellow]⚠ Microphone access needed[/]")
+        console.print(f"  [dim]macOS will prompt for permission on first recording.[/]")
+        console.print(f"  [dim]If denied, go to: System Settings → Privacy & Security → Microphone[/]")
+        console.print(f"  [dim]and enable access for your terminal app (iTerm2 / Terminal).[/]")
+
+    console.print(f"\n  [green]✓ Voice input ready![/]")
+    console.print(f"  [dim]Hold Space while typing prompt to record, release to transcribe.[/]")
+    return True
+
+
+def record_audio(max_seconds=30):
+    """Record from mic until stopped. Returns path to WAV file."""
+    wav_path = tempfile.mktemp(suffix=".wav")
+    proc = subprocess.Popen(
+        [SOX_REC, "-q", "-r", "16000", "-c", "1", "-b", "16", wav_path,
+         "trim", "0", str(max_seconds)],
+        stdin=subprocess.PIPE,
+        stdout=subprocess.DEVNULL,
+        stderr=subprocess.DEVNULL,
+    )
+    return proc, wav_path
+
+
+def stop_recording(proc):
+    """Stop the recording process."""
+    try:
+        proc.send_signal(signal.SIGINT)
+        proc.wait(timeout=3)
+    except:
+        proc.kill()
+
+
+def transcribe(wav_path, language="auto"):
+    """Transcribe audio using whisper-cli (CPU-only, no GPU conflict)."""
+    whisper = shutil.which("whisper-cli") or WHISPER_BIN
+    if not os.path.exists(whisper):
+        return None, "whisper-cli not found"
+
+    model = str(WHISPER_MODEL)
+    if not os.path.exists(model):
+        return None, "whisper model not found"
+
+    try:
+        # Use Metal GPU if llama-server has headroom (base model ~200MB),
+        # otherwise fall back to CPU-only
+        gpu_flags = []
+        try:
+            # Check free GPU headroom
+            out = subprocess.run(["pgrep", "-f", "llama-server"], capture_output=True, text=True)
+            if out.stdout.strip():
+                # llama-server running — whisper base needs ~200MB, check if safe
+                # With 2.7GB headroom on 24GB Mac, Metal whisper is fine
+                gpu_flags = []  # let whisper use Metal (default)
+            else:
+                gpu_flags = []  # no conflict, use Metal
+        except:
+            gpu_flags = ["--no-gpu"]  # safe fallback
+
+        result = subprocess.run(
+            [whisper,
+             "--model", model,
+             "--language", language,
+             "--no-timestamps",
+             "--threads", "8",
+             "--file", wav_path] + gpu_flags,
+            capture_output=True, text=True, timeout=30
+        )
+
+        # Parse output — whisper-cli outputs text lines (with --no-timestamps, just raw text)
+        lines = []
+        detected_lang = None
+        for line in result.stderr.split("\n"):
+            if "auto-detected language:" in line:
+                detected_lang = line.split("auto-detected language:")[-1].strip().split()[0]
+
+        for line in result.stdout.split("\n"):
+            line = line.strip()
+            # Skip empty lines and metadata
+            if line and not line.startswith("[") and not line.startswith("whisper_"):
+                lines.append(line)
+
+        text = " ".join(lines).strip()
+        # Clean up common whisper artifacts
+        text = text.replace("(silence)", "").replace("[BLANK_AUDIO]", "").strip()
+
+        return text, detected_lang
+
+    except subprocess.TimeoutExpired:
+        return None, "transcription timed out"
+    except Exception as e:
+        return None, str(e)
+    finally:
+        if os.path.exists(wav_path):
+            try:
+                os.unlink(wav_path)
+            except:
+                pass

localcoder-0.1.0.dist-info/METADATA

@@ -0,0 +1,187 @@
+Metadata-Version: 2.4
+Name: localcoder
+Version: 0.1.0
+Summary: 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.
+Project-URL: Homepage, https://github.com/AnassKartit/localcoder
+Project-URL: Repository, https://github.com/AnassKartit/localcoder
+Author: Anass Kartit
+License-Expression: Apache-2.0
+License-File: LICENSE
+Keywords: agent,ai,coding,gemma4,llama.cpp,local,localcoder,ollama,qwen
+Classifier: Development Status :: 4 - Beta
+Classifier: Environment :: Console
+Classifier: Intended Audience :: Developers
+Classifier: License :: OSI Approved :: Apache Software License
+Classifier: Programming Language :: Python :: 3
+Classifier: Topic :: Software Development :: Code Generators
+Requires-Python: >=3.10
+Requires-Dist: huggingface-hub>=0.20
+Requires-Dist: prompt-toolkit>=3.0
+Requires-Dist: rich>=13.0
+Description-Content-Type: text/markdown
+
+# localcoder
+
+**The local coding CLI that does the obvious things nobody else does.**
+
+```bash
+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](https://worldpopulationreview.com/country-rankings/cost-of-electricity-by-country). API: [anthropic.com](https://www.anthropic.com/pricing).*
+
+**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
+
+```bash
+# macOS (Apple Silicon)
+pipx install localcoder
+
+# First run — auto-detects hardware, shows what fits, starts model
+localcoder
+```
+
+Needs [llama.cpp](https://github.com/ggml-org/llama.cpp) or [Ollama](https://ollama.com). First run wizard handles this.
+
+## Commands
+
+```bash
+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:
+
+```bash
+pip install localfit
+localfit --launch claude --model gemma4-26b
+```
+
+One command: starts model → configures Claude Code → launches with `--bare` flag.
+See [localfit](https://github.com/AnassKartit/localfit) for details.
+
+### GPU Toolkit (localfit inside)
+
+```bash
+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 |
+
+```bash
+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
+
+```bash
+pip install pytest
+pytest tests/ -v      # 19 tests
+```

localcoder-0.1.0.dist-info/RECORD

@@ -0,0 +1,15 @@
+localcoder/__init__.py,sha256=eD-uC7S1gz0HbdaaORS2soj6zh5BB_tytoArAKfoioI,111
+localcoder/__main__.py,sha256=vFZX_R8P6RBS2CiBSvkFy_IG9b7wOi8d9h5H-y9Eskw,39
+localcoder/agent.py,sha256=tLaWfkgAmq-dQ9PqO59RziPS7deijg3HcJ1jmAyqWnM,1044
+localcoder/backends.py,sha256=JeGS-mzIWPElWmrZnswotInC225YhcFkRNztNI-3mhE,96500
+localcoder/bench.py,sha256=xVDRENdCrjy_Mz2Q_PeEax8ws_-6uTKcKYXmBm1eqr4,12620
+localcoder/cli.py,sha256=Yj6s02gQi37ctWydNt8VJHG0Zl1YdcZ4tM828oXWEeM,38917
+localcoder/gemma4coder_display.py,sha256=gGnVqmSpurOoBw_cup-ABWC_SzetWv1rk1KBz23fPck,17693
+localcoder/setup.py,sha256=3ACjQSOSmgPM9qhd5eaPAQ5xmGdYSB5C1y3EavRiw8M,11464
+localcoder/tui.py,sha256=wwOV8IrZumgg4QF1MCafHPtoCc5Av6sBoQoAqYnhoYU,9021
+localcoder/voice.py,sha256=ZEBFWzSB0tFoN8OQ68sV_jPPT_uMXcn95HoCS28ZmWw,7246
+localcoder-0.1.0.dist-info/METADATA,sha256=FOdoZ3sHADRJShBvD6-TaQO1hsQM5u-j3K-25_Rb9P0,6684
+localcoder-0.1.0.dist-info/WHEEL,sha256=QccIxa26bgl1E6uMy58deGWi-0aeIkkangHcxk2kWfw,87
+localcoder-0.1.0.dist-info/entry_points.txt,sha256=svs7yRCAJO8U8945pL-hic2eQnMfqRQx8v_m9U5J5Xo,51
+localcoder-0.1.0.dist-info/licenses/LICENSE,sha256=7KCxSOg3vOVwKw_kVsnzWCY7y0FNv4X_a0r__G0Q6hY,95
+localcoder-0.1.0.dist-info/RECORD,,

localcoder-0.1.0.dist-info/WHEEL

@@ -0,0 +1,4 @@
+Wheel-Version: 1.0
+Generator: hatchling 1.29.0
+Root-Is-Purelib: true
+Tag: py3-none-any

localcoder-0.1.0.dist-info/entry_points.txt

@@ -0,0 +1,2 @@
+[console_scripts]
+localcoder = localcoder.cli:main

localcoder-0.1.0.dist-info/licenses/LICENSE

@@ -0,0 +1,4 @@
+Apache License 2.0
+Copyright 2026 Anass Kartit
+
+Licensed under the Apache License, Version 2.0